{
  "updated": "2026-11-16",
  "count": 163,
  "license_note": "Free to use and quote with a link to https://www.shippedbyagents.dev. Summaries and labels are ours; numbers and quotes belong to the linked sources. source_type is vendor (published by the tool maker), builder (the team or person who did the work) or independent (third-party research). date is the source publish date (YYYY, YYYY-MM or YYYY-MM-DD), not when we added the story; null if the source gave none. Newest-first order and Latest story use this date. updated (YYYY-MM-DD) is the day we last changed the story data, with update_note saying what changed; null if never changed since publishing.",
  "stories": [
    {
      "slug": "keboola-zend-to-symfony-claude-code-two-sentence-prompt-13-prs-merged-22-hours",
      "title": "Keboola: a two-sentence prompt to Claude Code merges 13 Zend-removal PRs in about 22 hours through an existing agent harness",
      "url": "https://www.shippedbyagents.dev/stories/keboola-zend-to-symfony-claude-code-two-sentence-prompt-13-prs-merged-22-hours/",
      "source_url": "https://vojta.biberle.cz/en/blog/dark-factory-in-two-sentences",
      "source_name": "Vojta Biberle (blog)",
      "source_type": "builder",
      "date": "2026-11-12",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Merged PRs",
          "value": "13 in about 22 hours (3 more open)"
        },
        {
          "label": "Median open to merge",
          "value": "About 2 hours (fastest 57 minutes)"
        },
        {
          "label": "Dependencies removed",
          "value": "56 dependencies of the new codebase on the old one"
        },
        {
          "label": "Prompt",
          "value": "Two sentences"
        },
        {
          "label": "Remaining work",
          "value": "About 70 actions that need a human"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 13,
      "claimed_benefits": [
        {
          "claim": "13 PRs merged in about 22 hours from a two-sentence prompt",
          "metric": "13 PRs in ~22 hours"
        },
        {
          "claim": "Median open-to-merge time about two hours",
          "metric": "~2 hours"
        }
      ],
      "techniques": [
        "organiser session with subagents",
        "one agent per git worktree",
        "dependency map checked by a script",
        "risk-based merge gates with bot approval",
        "tasks from linear to merge queue"
      ],
      "caveats": "Written by the engineer who ran the experiment, on his personal blog; he does not state his role at Keboola. The merged changes were small. The harness was built beforehand and the post does not say how long that took. The 56-dependency figure appears in the post without saying whether it was all from this run. Overnight PRs waited 10–13 hours for review, and the E2E queue and human approvals were the main slowdowns. One task duplicated already-merged work after the agent read a stale local main.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "christopher-weir-ringleader-thread-90-claude-code-workers-60-merged-prs-five-weeks",
      "title": "Christopher Weir coordinates about 90 Claude Code worker sessions from one \"ringleader\" thread for more than 60 merged PRs in five weeks",
      "url": "https://www.shippedbyagents.dev/stories/christopher-weir-ringleader-thread-90-claude-code-workers-60-merged-prs-five-weeks/",
      "source_url": "https://christopher-weir.com/2026/11/coordinating-coding-agents-in-a-large-codebase",
      "source_name": "Christopher Weir (blog)",
      "source_type": "builder",
      "date": "2026-11-11",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Worker sessions",
          "value": "About 90 Claude Code sessions"
        },
        {
          "label": "Merged pull requests",
          "value": "More than 60"
        },
        {
          "label": "Duration",
          "value": "About five weeks"
        },
        {
          "label": "Human messages",
          "value": "About 200, typically around 15 words"
        },
        {
          "label": "Share of model output in workers",
          "value": "About nine-tenths (transcript estimate)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 60 merged PRs from about 90 worker sessions over about five weeks, steered with about 200 short messages",
          "metric": "60+ PRs, ~200 messages"
        },
        {
          "claim": "Hands-on checks by the ringleader caught a double-submit bug and wrong keyboard focus after hundreds of tests had passed"
        }
      ],
      "techniques": [
        "ringleader parent thread",
        "decisions written into the brief",
        "git worktree per worker",
        "parent reruns checks on combined code",
        "claude code skill"
      ],
      "caveats": "Self-reported by the developer about a work codebase; the company, product, stack and the kind of change behind the 60+ PRs are not named, so the work type is uncertain (the only named project is a later navigation refactor). Figures are rounded on purpose. He has not measured total cost or elapsed time against doing the project in one thread. In one case parallel workers filled the disk and stopped the local container runtime.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-11-11 from the page's time datetime=\"2026-11-11\"."
    },
    {
      "slug": "quera-cursor-claude-code-111-merged-prs-19-repos-empty-repo-to-merged-service-six-days",
      "title": "QuEra: one engineer with Cursor and Claude Code merges 111 PRs across 19 repos, empty repo to merged service in six days",
      "url": "https://www.shippedbyagents.dev/stories/quera-cursor-claude-code-111-merged-prs-19-repos-empty-repo-to-merged-service-six-days/",
      "source_url": "https://www.quera.com/blog-posts/what-compilers-never-needed-lessons-from-running-ai-agents-at-quera",
      "source_name": "QuEra blog",
      "source_type": "builder",
      "date": "2026-11-10",
      "tools": [
        "Cursor",
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Merged pull requests",
          "value": "111 across 19 repositories"
        },
        {
          "label": "Completed tasks",
          "value": "101"
        },
        {
          "label": "Code added",
          "value": "~11,200 lines production, ~25,400 lines tests"
        },
        {
          "label": "Empty repo to merged service (task profiles)",
          "value": "6 days (running on testbed and real QPU in 7)"
        },
        {
          "label": "User groups to all four clusters",
          "value": "88 days, engineer also on other projects"
        },
        {
          "label": "AI cost",
          "value": "~$3,900 across both projects"
        },
        {
          "label": "Internal estimate for same scope",
          "value": "9.5–13.5 engineer-months ($144,000–$205,000 salary), ±50%"
        },
        {
          "label": "Failed reviewer run",
          "value": "~500M cached-input tokens, ~$1,000"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Task profiles went from an empty repository to a merged service in six days",
          "metric": "6 days",
          "quote": "Task profiles went from an empty repository to a merged service in six days"
        },
        {
          "claim": "111 PRs merged across 19 repositories, adding about 11,200 lines of production code and 25,400 lines of tests",
          "metric": "111 merged PRs"
        },
        {
          "claim": "AI costs were about $3,900 against an estimated 9.5–13.5 engineer-months for the same scope",
          "metric": "~$3,900 vs 9.5–13.5 engineer-months",
          "quote": "That points to significant potential savings, but the comparison is incomplete."
        }
      ],
      "techniques": [
        "protected tests agents cannot change",
        "separate reviewer agents check against requirements",
        "nightly testing agent probes running service",
        "limits on spend, runtime and reads per run",
        "plan checklists and ai review records"
      ],
      "caveats": "Self-reported by QuEra. The 111-PR results come from interactive Cursor and Claude Code sessions by one engineer, not from unattended Quantopus runs, and the post keeps the two apart. QuEra itself says the $3,900 excludes the engineer's time and infrastructure, prices Claude at API rates rather than the actual subscription and misses earlier usage, that the effort estimate has about ±50% uncertainty, that there is no comparable project to measure speedup, and that test results were not independently rerun. All 111 PRs were merged by the engineer; 54 also had peer review. No author is named.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "fergus-6-3x-weekly-merges-same-team-pr-merge-time-18-hours-to-1-6",
      "title": "Fergus ships 6.3× more weekly merges with the same-sized team, cutting PR merge time from 18 hours to 1.6",
      "url": "https://www.shippedbyagents.dev/stories/fergus-6-3x-weekly-merges-same-team-pr-merge-time-18-hours-to-1-6/",
      "source_url": "https://www.linkedin.com/pulse/harder-better-faster-stronger-ben-gracewood-d2l9e",
      "source_name": "Ben Gracewood (Fergus)",
      "source_type": "builder",
      "date": "2026-11-09",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Weekly code merges",
          "value": "6.3×, same sized Fergus Engineering team"
        },
        {
          "label": "Time to merge PRs",
          "value": "18 hours → 1.6 hours"
        },
        {
          "label": "Engineering throughput",
          "value": "Up 500%"
        },
        {
          "label": "Monolith deploys",
          "value": "Up 250%"
        },
        {
          "label": "PRs per author",
          "value": "4.7×"
        },
        {
          "label": "App performance",
          "value": "Tripled"
        },
        {
          "label": "Change failure rate",
          "value": "Halved"
        },
        {
          "label": "Team size",
          "value": "Same as 18 months ago"
        },
        {
          "label": "Repositories",
          "value": "About 150 GitHub repos, including a core monolith"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "6.3× overall code merges per week with the same sized team",
          "metric": "6.3× weekly merges",
          "quote": "6.3x overall code merges per week, with the same sized Fergus Engineering team."
        },
        {
          "claim": "Time to merge PRs fell from 18 hours to 1.6 hours",
          "metric": "18 hours → 1.6 hours",
          "quote": "Our time to merge PRs has plummeted from 18 hours to 1.6 hours, mainly driven by AI reviews and more ownership over smaller services."
        },
        {
          "claim": "Engineering throughput up 500%",
          "metric": "Up 500%",
          "quote": "Engineering throughput up 500%, merges are 10 times faster, app performance tripled, and change failure rate halved."
        },
        {
          "claim": "Change failure rate halved, despite a handful of outages",
          "metric": "Change failure rate halved",
          "quote": "We have caused a handful of outages, but our change failure rate has halved."
        },
        {
          "claim": "4.7× pull requests per author",
          "metric": "4.7× PRs per author",
          "quote": "6.3x the PR rate overall, 4.7x PRs per author."
        }
      ],
      "techniques": [
        "ai reviews",
        "more ownership over smaller services",
        "pointing agents at heaviest endpoints",
        "offloading heavy queries",
        "modernise legacy stack while shipping features"
      ],
      "caveats": "Self-reported by Fergus's engineering lead. The post does not name which coding agents or AI review tools they used. Throughput (up 500%) and weekly merges (6.3×) are related but separately stated; raw PR counts are not given. 6.3× overall vs 4.7× per author is explained as more merged PRs coming from outside the engineering team. They report a handful of outages even as change failure rate halved. Comparison window for the 6.3× figure is not dated beyond the same team size as 18 months ago.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "alexander-marquardt-ecommerce-search-1652-merged-prs-a-month-coding-agent-fleet-0-2-percent-reverted",
      "title": "Alexander Marquardt runs a coding-agent fleet that merged 1,652 pull requests in a month on an e-commerce search solution, with 0.2 to 0.7% later reverted",
      "url": "https://www.shippedbyagents.dev/stories/alexander-marquardt-ecommerce-search-1652-merged-prs-a-month-coding-agent-fleet-0-2-percent-reverted/",
      "source_url": "https://alexmarquardt.com/ai-tools/coding-agents-codebase-under-control/",
      "source_name": "alexmarquardt.com (Alexander Marquardt)",
      "source_type": "builder",
      "date": "2026-11-08",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "mixed",
      "key_numbers": [
        {
          "label": "Merged PRs, main repo, October 2026",
          "value": "1,652 (about 373 a week), median 492 changed lines"
        },
        {
          "label": "Merged PRs, all four repos",
          "value": "2,760 in ISO weeks 41 to 44 (about 690 a week)"
        },
        {
          "label": "Merged PRs, main repo, April to 8 November 2026",
          "value": "6,234"
        },
        {
          "label": "Solution code",
          "value": "About 620,000 lines (backend and frontend)"
        },
        {
          "label": "Tests",
          "value": "2.1 million lines; 27,635 backend test functions and 12,467 frontend test cases"
        },
        {
          "label": "CI",
          "value": "58 workflows, 136 jobs"
        },
        {
          "label": "Merges later reverted",
          "value": "0.2 to 0.7%"
        },
        {
          "label": "Token efficiency",
          "value": "3.3x fewer tokens per changed line; 4.4x more PRs merged a week on about the same spend"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "1,652 pull requests merged in the main repository in October 2026",
          "metric": "1,652 merged PRs"
        },
        {
          "claim": "Only 0.2 to 0.7% of merges later reverted",
          "metric": "0.2 to 0.7%"
        },
        {
          "claim": "Each changed line takes less than a third of the tokens it did, with 4.4 times as many PRs merged a week on about the same spend",
          "metric": "3.3x drop in tokens per changed line; 4.4x PRs a week"
        }
      ],
      "techniques": [
        "adversarial review agent on every pr",
        "deterministic tests on recorded data",
        "every rule cites its incident",
        "control plane allocating agent slots by lane",
        "one clone per agent enforced by hooks"
      ],
      "caveats": "Self-reported by the engineer who runs the project, on his personal site; the employer and team size are not stated. Merged PR counts and line counts are throughput measures the author himself says are imperfect. The author says rework is about one fix round per merged PR, and he does not yet measure how many real shopper queries are interpreted correctly. Claude Code is named only in connection with hooks; the full set of agents and models is not stated.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "operum-two-founders-six-ai-agent-roles-296-merged-prs-12-days",
      "title": "Operum: two founders running six-role AI agent teams merge 296 PRs in 12 days, each gated on an AI tester approving the exact commit",
      "url": "https://www.shippedbyagents.dev/stories/operum-two-founders-six-ai-agent-roles-296-merged-prs-12-days/",
      "source_url": "https://dev.to/alprimak/two-people-six-ai-agents-296-merged-prs-in-12-days-g7m",
      "source_name": "DEV Community (Aleksandr Primak, Operum co-founder)",
      "source_type": "builder",
      "date": "2026-11-06",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Merged PRs",
          "value": "296 in 12 days (15–26 Oct 2026)"
        },
        {
          "label": "Merge rate",
          "value": "~25 a day, ~12 per person per day"
        },
        {
          "label": "PRs gated on CI plus AI Tester approval of exact commit",
          "value": "280 of 296"
        },
        {
          "label": "Mix by title",
          "value": "~49% user-visible, ~21% non-visible runtime, ~27% internal process, ~4% marketing"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "296 pull requests merged in 12 days by two people running AI agent teams",
          "metric": "296 PRs in 12 days"
        },
        {
          "claim": "About half of merged PRs were user-visible product changes",
          "metric": "~49% user-visible"
        }
      ],
      "techniques": [
        "role-based agent team via github issues and prs",
        "ai tester approves the exact commit",
        "approval resets on new push",
        "tester lists what no test covers",
        "auto-merge per team"
      ],
      "caveats": "Written by a co-founder about building Operum with Operum, in a post leading into its Product Hunt launch, so it promotes the product. The PR categories come from titles, not a diff audit, and the author says the PRs are small and PR count is the wrong scoreboard. Cursor is named as the app's default engine and recommendation; the post does not say which engine the team used for these 296 PRs, so the tool is Unstated. Revert and rework rates are not reported.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "ticksly-three-people-thirty-claude-code-codex-agents-2540-merge-requests-28-days",
      "title": "Ticksly, a three-person concert ticketing startup, has agents write all its code: 2,540 merge requests in 28 days at about $8 each",
      "url": "https://www.shippedbyagents.dev/stories/ticksly-three-people-thirty-claude-code-codex-agents-2540-merge-requests-28-days/",
      "source_url": "https://dev.to/dev1alone/how-we-run-thirty-claude-code-and-codex-agents-on-our-own-servers-and-the-tool-is-now-open-source-5c4e",
      "source_name": "Roman Kalmykov (Ticksly), DEV Community",
      "source_type": "builder",
      "date": "2026-11-05",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Merge requests",
          "value": "2,540 opened in the last 28 days"
        },
        {
          "label": "Cost",
          "value": "About $8 per merge request at API prices"
        },
        {
          "label": "Team",
          "value": "Three people; up to thirty agent sessions at once"
        },
        {
          "label": "Rewrite",
          "value": "harnessly rewritten from Node to Go in 19 days"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Agents opened 2,540 merge requests in 28 days",
          "metric": "2,540 merge requests in 28 days",
          "quote": "In the last 28 days they opened 2,540 merge requests"
        },
        {
          "claim": "About $8 per merge request at API prices",
          "metric": "about $8 each"
        },
        {
          "claim": "Agents rewrote harnessly from Node to Go in 19 days",
          "metric": "19 days"
        }
      ],
      "techniques": [
        "many parallel cli sessions in tmux",
        "shadow git repo per turn",
        "secrets kept out of agent chat",
        "per-person and per-model budgets",
        "agents calling agents"
      ],
      "caveats": "Written by the team that did the work, in a post launching their own open-source tool. Counts merge requests opened, not merged. No figures on quality, defects or revenue. Budgets were added after one week when subagents on an expensive model added $1,900 to the bill.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dropbox-new-experience-coding-agents-22-accessibility-features-53-test-cleanup-tasks-one-week",
      "title": "Dropbox engineers use coding agents to ship 22 accessibility features and clear a test cleanup deferred for two quarters in one week for the new Dropbox launch",
      "url": "https://www.shippedbyagents.dev/stories/dropbox-new-experience-coding-agents-22-accessibility-features-53-test-cleanup-tasks-one-week/",
      "source_url": "https://dropbox.tech/machine-learning/how-aiassisted-development-helped-us-build-the-new-dropbox-experience",
      "source_name": "Dropbox Tech blog (Pravi Garg, Sourabh Yerfule, Gennadiy Civil)",
      "source_type": "builder",
      "date": "2026-11-05",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Accessibility features",
          "value": "22 across Home, All Files, file previews and Photos"
        },
        {
          "label": "Test cleanup",
          "value": "53 tasks finished in one week, after being deferred for two quarters"
        },
        {
          "label": "New automated tests",
          "value": "Added for five core parts of the web app"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Cleanup deferred for two quarters finished in one week as 53 agent tasks",
          "metric": "53 tasks in one week"
        },
        {
          "claim": "22 accessibility features delivered before launch",
          "metric": "22 features"
        }
      ],
      "techniques": [
        "written plan agreed by engineer and agent",
        "agent runs builds and tests and records results",
        "instructions stored with the code",
        "cleanup split into small agent tasks",
        "agents try out the experience during development"
      ],
      "caveats": "Written by Dropbox engineers about their own launch. The post does not name the coding agents or models, and gives no counts of PRs, lines or time saved beyond the cleanup week. Some work described (tracing dependencies, writing adapters) is AI-assisted rather than delegated to agents.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "duolingo-ai-str-fix-claude-agent-sdk-investigates-and-fixes-shake-to-report-bugs",
      "title": "Duolingo builds AI STR Fix on the Claude Agent SDK to investigate and fix shake-to-report bugs, including release blockers",
      "url": "https://www.shippedbyagents.dev/stories/duolingo-ai-str-fix-claude-agent-sdk-investigates-and-fixes-shake-to-report-bugs/",
      "source_url": "https://blog.duolingo.com/ai-bug-fixing-agent/",
      "source_name": "Duolingo Blog (Shelby Wengreen)",
      "source_type": "builder",
      "date": "2026-11-04",
      "tools": [
        "Claude Agent SDK"
      ],
      "models": [
        "Claude Sonnet",
        "Claude Opus"
      ],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Outcome",
          "value": "Has fixed release-blocking bugs"
        },
        {
          "label": "Time saved",
          "value": "Automates hours of manual work on each STR bug ticket it runs on"
        },
        {
          "label": "Company size",
          "value": "Nearly 1,000 Duos"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Has fixed release-blocking bugs",
          "quote": "AI STR Fix has fixed release-blocking bugs."
        },
        {
          "claim": "Automates hours of manual work per bug ticket",
          "metric": "Hours of manual work on each STR bug ticket",
          "quote": "It automates hours of manual work on each STR bug ticket it runs on."
        }
      ],
      "techniques": [
        "document findings as files in the working directory",
        "heuristic pre-filter before the agent",
        "self-review step",
        "fix ci stage",
        "feature research before implementation"
      ],
      "caveats": "Written by Duolingo's DevX AI team about its own tool. No counts are given for tickets processed, fixes merged, success rate or cost. \"Hours of manual work\" per ticket is stated without a measured figure. Platforms are implied (Android and iOS experiments) rather than listed as a stack.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "bloomberg-automated-uplift-hybrid-llm-pipelines-2333-prs-729-migrations",
      "title": "Bloomberg's Automated Uplift ran LLM-assisted migrations across thousands of repositories: 2,333 PRs and 729 successful migrations",
      "url": "https://www.shippedbyagents.dev/stories/bloomberg-automated-uplift-hybrid-llm-pipelines-2333-prs-729-migrations/",
      "source_url": "https://arxiv.org/abs/2611.05457",
      "source_name": "Corchero et al., arXiv:2611.05457",
      "source_type": "builder",
      "date": "2026-11-03",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Pull requests",
          "value": "2,333 generated across two campaigns"
        },
        {
          "label": "Successful migrations",
          "value": "729"
        },
        {
          "label": "Python campaign",
          "value": "About 1,900 candidate repositories"
        },
        {
          "label": "C++ campaign",
          "value": "About 6,700 candidate repositories"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Two campaigns produced 2,333 pull requests and 729 successful migrations",
          "metric": "2,333 pull requests, 729 successful migrations",
          "quote": "Together, these campaigns generated 2,333 pull requests, resulting in 729 successful migrations."
        }
      ],
      "techniques": [
        "deterministic pipeline with bounded llm steps",
        "pre-computed context",
        "production telemetry as context",
        "verify in pipeline before opening pr"
      ],
      "caveats": "Written mostly by Bloomberg engineers who ran the work, with an academic co-author. Drafted from the abstract only; the abstract names no model or coding tool and gives no time or cost figures. 729 successes out of 2,333 PRs means many PRs did not result in a migration.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "microsoft-copilot-cowork-nine-engineers-1211-prs-python-to-rust",
      "title": "Microsoft's Copilot Cowork team goes empty-repo to worldwide GA in 4.5 months with agents writing most of the code, then ports the live runtime from Python to Rust",
      "url": "https://www.shippedbyagents.dev/stories/microsoft-copilot-cowork-nine-engineers-1211-prs-python-to-rust/",
      "source_url": "https://commandline.microsoft.com/how-we-built-copilot-cowork-python-to-rust-migration/",
      "source_name": "Microsoft Command Line blog (Bas Brekelmans)",
      "source_type": "builder",
      "date": "2026-11-02",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "mixed",
      "key_numbers": [
        {
          "label": "Empty repo to worldwide GA",
          "value": "13 Jan to 16 Jun 2026 (~4.5 months)"
        },
        {
          "label": "Merged PRs per month",
          "value": "257 in January → 2,754 in May; busiest pre-GA week 783"
        },
        {
          "label": "Peak day",
          "value": "1,211 PRs merged by nine engineers on 21 Oct 2026"
        },
        {
          "label": "Factory week",
          "value": "4,581 PRs merged week of 12 Oct (~650/day)"
        },
        {
          "label": "Python deleted at cutover",
          "value": "~8,000 files and 4.9 million lines (incl. tests and generated code) on 9 Oct"
        },
        {
          "label": "First-token latency (same 12h window)",
          "value": "Python p50 7.9s / p90 14.5s → Rust p50 2.7s / p90 6.2s"
        },
        {
          "label": "Parity fixtures",
          "value": "~430 fixtures generated from live Python behaviour"
        },
        {
          "label": "Test-speed harness",
          "value": "~300 merged PRs; touched tests ~17,700s → ~1,700s CI time"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Worldwide GA about four and a half months after the first commit",
          "metric": "~4.5 months",
          "quote": "We reached worldwide general availability on June 16, about four and a half months after the first commit"
        },
        {
          "claim": "Time to first token on the first turn of a session dropped a lot after the Rust cutover",
          "metric": "p50 7.9s → 2.7s",
          "quote": "Time to first token on the first turn of a session dropped a lot."
        },
        {
          "claim": "Touched test files cut CI time from roughly 17,700 seconds to about 1,700 across ~300 harness PRs",
          "metric": "~17,700s → ~1,700s",
          "quote": "Across them, the touched test files went from roughly 17,700 seconds of CI time to about 1,700."
        }
      ],
      "techniques": [
        "markdown spec before code",
        "narrow parallel agents",
        "competing agent reviewers",
        "checks that fail on purpose",
        "python behaviour fixtures for rust parity",
        "automated factory loops with attempt limits"
      ],
      "caveats": "Self-reported Microsoft eng post. The specific coding-agent product is not named (only “agents”). Latency gain partly attributed to moving the agent loop out of the container, not Rust alone. No cost-per-turn claim.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "harvard-jellyfish-ai-agents-review-bottleneck",
      "title": "Harvard study of 655 firms: AI agents raise code volume far more than shipped work, review becomes the bottleneck",
      "url": "https://www.shippedbyagents.dev/stories/harvard-jellyfish-ai-agents-review-bottleneck/",
      "source_url": "https://fion.ac/jellyfish.pdf",
      "source_name": "Chen & Stratton, Harvard (job market paper); coverage in Ars Technica",
      "source_type": "independent",
      "date": "2026-11-01",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Firms",
          "value": "655"
        },
        {
          "label": "Workers",
          "value": "720,542"
        },
        {
          "label": "Work events",
          "value": "Over 445 million (Jan 2021–Aug 2026)"
        },
        {
          "label": "Lines of code, 12 months after agent adoption",
          "value": "+44.0% (marginally insignificant)"
        },
        {
          "label": "Commits",
          "value": "+36.0%"
        },
        {
          "label": "Pull requests opened",
          "value": "+35.7%"
        },
        {
          "label": "Pull requests merged",
          "value": "+26.0%"
        },
        {
          "label": "Resolved Jira issues / epics",
          "value": "+13.7% / +18.0% (both marginally insignificant)"
        },
        {
          "label": "Review time (PR submitted → merged)",
          "value": "+56.1%"
        },
        {
          "label": "Review comments per PR",
          "value": "+38.6%"
        },
        {
          "label": "Engineering employment",
          "value": "No significant effect (−2.6%)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Adopting AI coding tools substantially increases code output, though the effect on final software output is much smaller",
          "quote": "Adoption substantially increases code output"
        },
        {
          "claim": "AI coding agents raised lines of code, commits and pull requests opened",
          "metric": "+44.0% / +36.0% / +35.7%",
          "quote": "44.0% for lines of code, 36.0% for commits, and 35.7% for pull requests opened"
        }
      ],
      "techniques": [
        "staggered difference-in-differences",
        "code review as the bottleneck"
      ],
      "caveats": "Current version dated 1 Nov 2026 (first version 7 Jan 2026). The earlier version we first used covered 718 firms and ~300M events. Data from Jellyfish, an engineering analytics vendor; the authors state the conclusions are their own and that Jellyfish reviewed the paper only for proprietary information. Effects are firm-level, so they average over workers who did and didn't use the tools. The Ars Technica coverage describes the earlier version.",
      "updated": "2026-11-02",
      "update_note": "Now uses the newer version of the study (1 Nov 2026): 655 firms and 445M+ events, previously 718 firms and ~300M. All numbers now come from the paper itself."
    },
    {
      "slug": "nicam-climate-model-fortran-to-kokkos-claude-code-140000-lines-to-35000-1-71x-faster",
      "title": "University of Tokyo researchers port the NICAM climate model from 140,000 lines of Fortran to 35,000 lines of C++/Kokkos with Claude Code, running 1.71x faster on GH200",
      "url": "https://www.shippedbyagents.dev/stories/nicam-climate-model-fortran-to-kokkos-claude-code-140000-lines-to-35000-1-71x-faster/",
      "source_url": "https://arxiv.org/abs/2610.41313",
      "source_name": "arXiv (Ren, Shimokawabe, Yashiro)",
      "source_type": "builder",
      "date": "2026-10-30",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Fable 5"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Code ported",
          "value": "~140,000 lines of Fortran analysed and ported into ~35,000 lines of C++/Kokkos"
        },
        {
          "label": "Scope",
          "value": "Dynamical core and seven physics packages"
        },
        {
          "label": "Application speedup vs Fortran/OpenACC on GH200",
          "value": "1.71x on GH200, 1.54x on MI300A"
        },
        {
          "label": "Active workflow days",
          "value": "23"
        },
        {
          "label": "Human instructions",
          "value": "797 (6,007 assistant messages)"
        },
        {
          "label": "Token cost",
          "value": "~USD 5,589 estimated at API list prices (used under a subscription plan)"
        },
        {
          "label": "Ensemble agreement",
          "value": "95.3% of 5,900 member–quantity scores within |z| ≤ 2 in both implementations"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Port runs 1.71x faster on GH200 and 1.54x on MI300A than the Fortran/OpenACC baseline",
          "metric": "1.71x / 1.54x"
        },
        {
          "claim": "Port is statistically consistent with the original Fortran model in 100-member ensemble tests"
        },
        {
          "claim": "One codebase now runs on both NVIDIA and AMD GPUs through Kokkos CUDA and HIP backends"
        }
      ],
      "techniques": [
        "kickoff document with rules",
        "plan mode before coding",
        "module-by-module milestones",
        "reference dumps from fortran as unit tests",
        "human review at each milestone"
      ],
      "caveats": "Written by the researchers who did the port. The cost is an estimate at API list prices; they actually used a subscription plan. Human working hours were not recorded (instruction counts come from Claude Code's own logs), and days spent only on human review or cluster work are not in the 23-day count. Branches outside the selected configuration were not ported, and the authors credit NICAM's consistent, well-maintained Fortran for part of the smooth run. Human review and the subscription's five-hour usage limits were the main constraints.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "spawn-ten-person-team-agent-fleet-77000-changes-30-days-3-1m-tokens",
      "title": "Spawn, a ten-person game engine startup, says its agent fleet shipped 77,000 substantial changes in 30 days on $3.1M of tokens",
      "url": "https://www.shippedbyagents.dev/stories/spawn-ten-person-team-agent-fleet-77000-changes-30-days-3-1m-tokens/",
      "source_url": "https://x.com/jsnnsa/status/2115623382721126826",
      "source_name": "X Article by Spawn's founder (@jsnnsa), \"How I Built the Most AI-Pilled Company in the World\"",
      "source_type": "builder",
      "date": "2026-10-29",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Fable"
      ],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Changes shipped",
          "value": "77,000 substantial changes in the last 30 days, about 108 an hour"
        },
        {
          "label": "Token spend",
          "value": "$3.1 million on tokens last month"
        },
        {
          "label": "Reverts",
          "value": "1,055 of 123,548 changes over the last year, about 0.9%"
        },
        {
          "label": "AI judges",
          "value": "More than 11,000 verdicts in eight days from October 6 (7,296 decided, 2,213 sent back, 1,863 held for a human)"
        },
        {
          "label": "Daily changes",
          "value": "About 1,000 a day the week before the judges, 7,250 on October 10"
        },
        {
          "label": "Renderer rewrite",
          "value": "Rewritten in six days, about 25x more efficient"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "77,000 substantial changes shipped in the last 30 days, about 108 an hour, and 500x more than six months earlier",
          "metric": "77,000 changes/30 days; 500x",
          "quote": "We now ship 500x more than we did 6 months ago."
        },
        {
          "claim": "Only about 0.9% of changes reverted over the last year",
          "metric": "0.9% reverts",
          "quote": "over the last year only 1,055 of our 123,548 changes were reverted, about 0.9%"
        }
      ],
      "techniques": [
        "permanent named agents that own an area",
        "swarms of helper agents",
        "stigmergy instead of coordination",
        "helper agents try to break each change",
        "ai judges"
      ],
      "caveats": "These are the founder's own claims about the company they run, unverified, published as an X Article rather than a blog post. No code, dashboards or outside data back the numbers. The post defines a substantial change as one touching engine code or game logic but doesn't show how changes are counted. The revert rate covers the last year, not the 30-day window. The post says the team is ten people but also refers to five people working alongside the agents. The author is named only by X handle and display name (jacob).",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dana-indonesia-android-design-system-migration-cursor-agents-notion-52-minutes",
      "title": "DANA Indonesia cut a design-system component migration from 24 hours to 52 minutes per batch with Cursor agents run from Notion",
      "url": "https://www.shippedbyagents.dev/stories/dana-indonesia-android-design-system-migration-cursor-agents-notion-52-minutes/",
      "source_url": "https://www.notion.com/customers/dana-indonesia-code-migration",
      "source_name": "Notion customer story",
      "source_type": "vendor",
      "date": "2026-10-28",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Time from assignment to review-ready PR per batch",
          "value": "52 minutes, down from an average of 24 hours"
        },
        {
          "label": "Components migrated so far",
          "value": "575 of 2,970, at 180 a week"
        },
        {
          "label": "Environment startup time",
          "value": "Down 47%, from 17 to 9 minutes"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Time from assignment to a review-ready PR cut from an average of 24 hours to 52 minutes per batch",
          "metric": "24 h → 52 min",
          "quote": "A batch now moves from assignment to a review-ready PR in 52 minutes, compared with a previous average of 24 hours"
        },
        {
          "claim": "575 components migrated so far, at 180 a week",
          "metric": "180/week"
        },
        {
          "claim": "Environment startup time down 47%, from 17 to 9 minutes",
          "metric": "-47%",
          "quote": "reduces average environment startup time by 47% (from 17 to 9 minutes)"
        }
      ],
      "techniques": [
        "backlog batched in notion",
        "notion external agent",
        "team of specialized cursor agents",
        "migration agent opens the branch"
      ],
      "caveats": "Vendor customer story from Notion, which sells the External Agent feature. Counts PRs reaching review, not merged; no figures for review time, rework or defects. The page says most customers use iPhones while the migration is the Android app.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-10-28 from the page's JSON-LD datePublished."
    },
    {
      "slug": "x2-labs-bzflag-browser-game-claude-code-overnight-seven-days-2421-tests",
      "title": "X2 Labs founder rebuilds BZFlag as a multiplayer browser game with Claude Code in seven days, mostly overnight, with 2,421 tests",
      "url": "https://www.shippedbyagents.dev/stories/x2-labs-bzflag-browser-game-claude-code-overnight-seven-days-2421-tests/",
      "source_url": "https://x2labs.ai/notes/bzflag-in-the-browser",
      "source_name": "X2 Labs notes (George Purusniuc)",
      "source_type": "builder",
      "date": "2026-10-28",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 5.5",
        "Claude Sonnet 5.5"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "7 days from first commit to public"
        },
        {
          "label": "Plans and tasks",
          "value": "6 plans, with 106 tasks"
        },
        {
          "label": "Commits",
          "value": "210, all co-authored by Claude; 167 between 6 p.m. and 8 a.m."
        },
        {
          "label": "Game code",
          "value": "About 43,000 lines of TypeScript"
        },
        {
          "label": "Test code",
          "value": "About 27,000 lines"
        },
        {
          "label": "Automated tests",
          "value": "2,421, plus 7 browser test files"
        },
        {
          "label": "Ported from C++",
          "value": "55 files from BZFlag"
        }
      ],
      "duration_days": 7,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Public multiplayer game seven days after the first commit",
          "metric": "7 days",
          "quote": "a playable game, public seven days after the first commit, with 2,421 automated tests."
        },
        {
          "claim": "2,421 automated tests written with every task",
          "metric": "2,421 tests",
          "quote": "I would never write 2,421 tests myself."
        }
      ],
      "techniques": [
        "plans by day, building by night",
        "plans of around 20 small tasks",
        "helper agent builds, second agent reviews",
        "every task must pass the tests",
        "original source code as the spec"
      ],
      "caveats": "Self-reported by the builder on his own company's site. A hobby game, not a commercial product; the post does not report players, cost or token usage. Numbers are counted from the git history, plans and test runner on 28 October 2026. The game reuses BZFlag's rules and source as a specification.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "microsoft-intune-dotnet-8-migration-github-copilot-modernization-six-weeks-to-1-5",
      "title": "Microsoft Intune engineer moves an enrollment service library to .NET 8 in 1.5 weeks against a six-week estimate with GitHub Copilot modernization",
      "url": "https://www.shippedbyagents.dev/stories/microsoft-intune-dotnet-8-migration-github-copilot-modernization-six-weeks-to-1-5/",
      "source_url": "https://techcommunity.microsoft.com/blog/appsonazureblog/how-microsoft-intune-team-upgraded-a-net-service-with-github-copilot-modernizati/4566987",
      "source_name": "Microsoft Community Hub, Apps on Azure blog (RamyaSharma)",
      "source_type": "vendor",
      "date": "2026-10-27",
      "tools": [
        "GitHub Copilot"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "First library",
          "value": "Estimated at about six weeks, done in one and a half (75% schedule reduction)"
        },
        {
          "label": "Time saved",
          "value": "More than four weeks on the first library; a second six-week library also came in early (no figure given)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A library estimated at about six weeks was done in one and a half, a 75% schedule reduction",
          "metric": "6 weeks → 1.5 weeks (-75%)",
          "quote": "a 75% schedule reduction"
        },
        {
          "claim": "A second six-week library also came in early, with no figure given"
        }
      ],
      "techniques": [
        "copilot modernization for mechanical work",
        "three working files kept alongside the code",
        "project file conversion",
        "nuget package resolution"
      ],
      "caveats": "Microsoft writing about its own GitHub Copilot product in its Customer Zero series. Savings are against the team's own estimates. No line counts, defect or cost figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "merciyanis-platform-rewrite-250k-lines-10-repos-one-day-30-agents",
      "title": "MerciYanis rewrites its whole platform, about 250,000 lines across 10 repositories, in one day with one engineer and around 30 parallel agents",
      "url": "https://www.shippedbyagents.dev/stories/merciyanis-platform-rewrite-250k-lines-10-repos-one-day-30-agents/",
      "source_url": "https://engineering.merciyanis.com/blog/refactoring-codebases-at-scale",
      "source_name": "MerciYanis Engineering Blog (Matthieu Jabbour, CTO)",
      "source_type": "builder",
      "date": "2026-10-26",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Scope",
          "value": "Roughly 250,000 lines across 10 repositories"
        },
        {
          "label": "Time and people",
          "value": "One engineer's full-time day, about 30 agents in parallel"
        },
        {
          "label": "Tokens",
          "value": "About 1.6 billion"
        },
        {
          "label": "Bugs fixed along the way",
          "value": "About 50 bugs and edge cases"
        },
        {
          "label": "Regressions",
          "value": "Two small regressions found in QA, both fixed within minutes"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Roughly 250,000 lines across 10 repositories rewritten in a single day by one engineer and about 30 agents",
          "metric": "250k lines in 1 day",
          "quote": "All of it: roughly 250,000 lines of code, spread across 10 repositories, in a single day."
        },
        {
          "claim": "About 50 bugs and edge cases fixed along the way; two small regressions found in QA and fixed within minutes",
          "metric": "~50 bugs fixed; 2 regressions"
        }
      ],
      "techniques": [
        "100% coverage before rewriting",
        "tests only on public contracts",
        "markdown handbook as agent context",
        "about 30 agents in parallel",
        "no line-by-line review"
      ],
      "caveats": "Written by the company's own CTO. The one day excludes the earlier work raising test coverage from an average of about 60% to 100%, which the post says took most of the time. Results are reported from the QA environment; no production figures. Coding tool not named. Stack not stated; the post's code sample looks like TypeScript and its diagram names PostgreSQL and Kafka, so check before filling it in.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "michael-cutler-mylocker-gym-booking-app-claude-code-skill-to-app-store-codex-107-prs",
      "title": "A Claude Code skill for booking gym classes becomes MyLocker, an App Store and Google Play app now run day to day by Codex, with 107 merged PRs",
      "url": "https://www.shippedbyagents.dev/stories/michael-cutler-mylocker-gym-booking-app-claude-code-skill-to-app-store-codex-107-prs/",
      "source_url": "https://cutler.sg/blog/2026-10-gym-booking-skill-to-app-store",
      "source_name": "Fractional Futurist blog (Michael Cutler)",
      "source_type": "builder",
      "date": "2026-10-25",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Backend cutover",
          "value": "3 days"
        },
        {
          "label": "Web and native apps",
          "value": "Less than a day of building; over a week to get approved"
        },
        {
          "label": "Merged PRs",
          "value": "107 by 20 October 2026"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 107,
      "claimed_benefits": [
        {
          "claim": "Skill moved to the gym's new backend in three days, and the web app plus both native apps took less than a day of building",
          "metric": "3 days; under a day",
          "quote": "End to end, the web app and both native apps took less than a day of building"
        }
      ],
      "techniques": [
        "claude code skill first, app later",
        "verified on a real account",
        "codex runs the repo day to day",
        "correcting the agent's domain assumptions"
      ],
      "caveats": "Self-reported by the builder. Analytics are off, so there are no user numbers. Unofficial app for one gym chain's members.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "codemeepo-multi-agent-repo-migration-copilot-cli-18-percent-more-tests",
      "title": "CodeMeepo multi-agent migrations with Copilot CLI pass 18% more tests than ReCodeAgent across nine repos, including a full SONiC xcvrd Python-to-Rust port",
      "url": "https://www.shippedbyagents.dev/stories/codemeepo-multi-agent-repo-migration-copilot-cli-18-percent-more-tests/",
      "source_url": "https://arxiv.org/abs/2610.34633",
      "source_name": "arXiv 2610.34633 (Habibi, Wu, Jiang, Lu)",
      "source_type": "independent",
      "date": "2026-10-25",
      "tools": [
        "GitHub Copilot"
      ],
      "models": [
        "claude-sonnet-5"
      ],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Subjects",
          "value": "9 repos, 3 language pairs"
        },
        {
          "label": "Tests passed vs ReCodeAgent",
          "value": "1,334 vs 1,129 (+18%)"
        },
        {
          "label": "xcvrd (production daemon)",
          "value": "105/105 oracle tests vs ReCodeAgent 53/105"
        },
        {
          "label": "Optimization",
          "value": "5 of 10 xcvrd benchmarks improved >5%, avg 25% on those"
        },
        {
          "label": "Model (all arms)",
          "value": "claude-sonnet-5 via Copilot CLI"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Best pass rate on every subject; 18% more tests passed than ReCodeAgent on the same model",
          "metric": "+18% tests (1,334 vs 1,129)",
          "quote": "passes 18% more tests than a state-of-the-art multi-agent baseline using the same model"
        },
        {
          "claim": "Full 105/105 pass on production SONiC xcvrd Python→Rust vs 53 for ReCodeAgent",
          "metric": "105/105 vs 53/105"
        },
        {
          "claim": "Optimization phase improved half of xcvrd benchmarks by 25% on average",
          "metric": "5/10 benchmarks, avg +25%",
          "quote": "its optimization phase then improves half of the benchmarks, by 25% on average"
        }
      ],
      "techniques": [
        "fixed external acceptance oracle",
        "milestone cumulative test gates",
        "parity gap re-scoping loop",
        "file-artifact agent handoffs",
        "benchmark-guided optimization phase"
      ],
      "caveats": "Research paper; authors affiliated with UC Irvine, Microsoft, and OpenAI, with contributions while at Microsoft. Benchmarks are controlled migrations with fixed oracles, not live product ship stories.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-10-25 from the arXiv submission date on abs/2610.34633."
    },
    {
      "slug": "taewoo-tabby-tab-manager-claude-code-19-days-141-users-909-dollars",
      "title": "Developer builds and ships the Tabby tab-manager browser extension in 19 days by talking to Claude Code, reaching 141 users at about $909 in API-priced tokens",
      "url": "https://www.shippedbyagents.dev/stories/taewoo-tabby-tab-manager-claude-code-19-days-141-users-909-dollars/",
      "source_url": "https://dev.to/taewoo/19-days-546-messages-141-users-building-my-app-by-mostly-talking-to-claude-code-1mgb",
      "source_name": "DEV Community (taewoo)",
      "source_type": "builder",
      "date": "2026-10-24",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "19 days since 5 Oct; v1.2 about a day from plan to release"
        },
        {
          "label": "Messages",
          "value": "546 over the project, 58 for v1.2"
        },
        {
          "label": "Tokens",
          "value": "2.5B for the project (888M for v1.2)"
        },
        {
          "label": "Cost at API prices",
          "value": "About $909 for the project (about $316 for v1.2); actually paid a flat Claude Max plan"
        },
        {
          "label": "Users",
          "value": "141 on the Chrome and Firefox stores"
        }
      ],
      "duration_days": 19,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "141 users on the Chrome and Firefox stores 19 days after the first line of code, at about $6 of API-priced AI time per user",
          "metric": "141 users; ~$6/user",
          "quote": "141 users on the Chrome and Firefox stores, 19 days after the first line of code."
        }
      ],
      "techniques": [
        "talking to the agent in vietnamese",
        "approving plans and judging screenshots",
        "short demo clips as feedback",
        "each bug report becomes a test",
        "ask it to write the lesson down"
      ],
      "caveats": "Self-reported, and the post is marked AI-assisted on DEV. Costs are estimates from token logs at Opus list prices, not what he paid. Part of the work counted was promo videos, not code. Small user numbers.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "alex-cloudstar-site-rebuild-nextjs-16-claude-code-two-days-543-tests",
      "title": "Solo developer rebuilds his Astro site on Next.js 16 with Claude Code in two days: 228 posts reworked, 543 tests, no broken URLs",
      "url": "https://www.shippedbyagents.dev/stories/alex-cloudstar-site-rebuild-nextjs-16-claude-code-two-days-543-tests/",
      "source_url": "https://dev.to/alexcloudstar/i-shipped-32784-lines-of-code-in-2-days-with-claude-code-here-is-what-the-number-hides-276d",
      "source_name": "Alex Cloudstar (DEV Community, originally on alexcloudstar.com)",
      "source_type": "builder",
      "date": "2026-10-23",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Two days (late 21 Oct to morning of 23 Oct 2026)"
        },
        {
          "label": "Lines",
          "value": "32,784, including generated MDX content, tests and code later deleted"
        },
        {
          "label": "Commits and PRs",
          "value": "About 100 commits, 11 merged PRs"
        },
        {
          "label": "Tests",
          "value": "543 passing"
        },
        {
          "label": "Posts reworked",
          "value": "228"
        },
        {
          "label": "Broken URLs",
          "value": "Zero"
        }
      ],
      "duration_days": 2,
      "cost_usd": null,
      "prs_merged": 11,
      "claimed_benefits": [
        {
          "claim": "A full site rebuild on Next.js 16 in two days, with 228 posts reworked, 543 tests passing and zero broken URLs",
          "metric": "2 days; 543 tests; 0 broken URLs",
          "quote": "a full rebuild of this site on Next.js 16, 228 blog posts rewritten against real keyword data, 543 tests passing, and zero broken URLs"
        }
      ],
      "techniques": [
        "keep every merge small enough to review",
        "tests for metadata and sitemap",
        "posts reworked against keyword data"
      ],
      "caveats": "Author's own account of his own site; he writes mostly about Claude Code, so it's partly content marketing. A small personal site rather than production software for users. No cost or usage figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "parkease-solo-parking-marketplace-claude-code-five-weeks-86000-lines-4300-tests",
      "title": "Solo developer builds ParkEase, a five-role parking marketplace with about 86,000 lines of TypeScript and 4,300+ tests, in five weeks with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/parkease-solo-parking-marketplace-claude-code-five-weeks-86000-lines-4300-tests/",
      "source_url": "https://dev.to/rakno/i-built-a-parking-marketplace-in-five-weeks-the-ai-wrote-most-of-the-code-i-did-the-other-part-3io",
      "source_name": "Onkar Deokate (DEV Community)",
      "source_type": "builder",
      "date": "2026-10-23",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Five weeks (Sep 10 – Oct 13, 2026 per the repo)"
        },
        {
          "label": "Merged PRs",
          "value": "29 (the repo README says 32)"
        },
        {
          "label": "Code",
          "value": "~86,000 lines of TypeScript across 5 apps and 6 packages"
        },
        {
          "label": "Tests",
          "value": "3,246 unit and 1,080 integration tests on real Postgres"
        },
        {
          "label": "Architecture decision records",
          "value": "34"
        },
        {
          "label": "Learnings file",
          "value": "246 entries"
        }
      ],
      "duration_days": 35,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Five-app marketplace with ~86,000 lines and 4,300+ tests built solo in five weeks",
          "metric": "5 weeks, ~86,000 lines"
        },
        {
          "claim": "Features he would normally cut for time, like an admin audit log, surge heat map and load tests, all fit"
        }
      ],
      "techniques": [
        "architecture decision records",
        "project rules file read every session",
        "agent renders design options to choose from",
        "learnings file read before every task",
        "fixed workflow with gates"
      ],
      "caveats": "Self-reported by the builder. The post and the repo README differ on merged PRs (29 vs 32). ParkEase is live as a demo; no real users, bookings or revenue are reported. No AI cost is given in this post.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "getyourguide-automerge-5000-dependency-prs-a-month-ai-ci-fixes",
      "title": "GetYourGuide's Automerge merges 5,000+ dependency PRs a month across 250+ repos, with an AI agent fixing broken CI",
      "url": "https://www.shippedbyagents.dev/stories/getyourguide-automerge-5000-dependency-prs-a-month-ai-ci-fixes/",
      "source_url": "https://www.getyourguide.careers/posts/from-backlog-to-autopilot-our-ai-workflow-that-merges-5-000-pull-requests-a-month",
      "source_name": "Inside GetYourGuide engineering blog",
      "source_type": "builder",
      "date": "2026-10-22",
      "tools": [],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Dependency PRs merged",
          "value": "\"more than 5,000\" a month with \"zero human clicks\""
        },
        {
          "label": "Repos",
          "value": "\"more than 250\", up from 4 in 2023"
        },
        {
          "label": "AI fixes",
          "value": "\"roughly 1,200 CI and security fixes a month\" (one month: 646 fixed with AI, 579 security issues fixed with AI)"
        },
        {
          "label": "Automation rate",
          "value": "About 97.2% of dependency PRs handled automatically, 2.8% by engineers"
        },
        {
          "label": "Time saved",
          "value": "\"roughly 830 engineer-hours a month\" (their back-of-envelope at 10 minutes per PR)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 5,000 dependency PRs merged a month across 250+ repos, with about 97.2% handled automatically",
          "metric": "5,000+ PRs/month; 97.2% automatic",
          "quote": "Automerge currently handles around 97,2% of dependency PRs automatically"
        },
        {
          "claim": "AI handles roughly 1,200 CI and security fixes a month",
          "metric": "~1,200 fixes/month",
          "quote": "AI handles roughly 1,200 CI and security fixes a month"
        },
        {
          "claim": "Back-of-envelope: handling 5,000 PRs by hand would take about 830 engineer-hours a month",
          "metric": "~830 hours/month",
          "quote": "handling 5,000 PRs manually would take roughly 830 engineer-hours a month"
        }
      ],
      "techniques": [
        "one-line per-repo merge policy",
        "agent fixes failing ci on dependency bumps",
        "agent resolves merge conflicts"
      ],
      "caveats": "Self-reported on GetYourGuide's careers blog. It doesn't name the AI agent or model. Most of the 5,000 merges are routine Dependabot bumps merged by rules; the AI part is the roughly 1,200 monthly fixes. The post admits it sometimes auto-merges faulty PRs but gives no count of bad merges or reverts. The hours saved figure is an estimate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dsalvado-shuttlescp-rust-gtk-sftp-client-claude-code-22k-lines-149-tests",
      "title": "A database specialist who doesn't read Rust ships ShuttleSCP, a 22,000-line Rust and GTK 4 file-transfer client, with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/dsalvado-shuttlescp-rust-gtk-sftp-client-claude-code-22k-lines-149-tests/",
      "source_url": "https://dev.to/dsalvado/what-i-learned-building-a-real-app-with-claude-code-1af",
      "source_name": "DEV Community (dsalvado)",
      "source_type": "builder",
      "date": "2026-10-22",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Size",
          "value": "About 22,000 lines of Rust, 149 automated tests"
        },
        {
          "label": "Releases",
          "value": "Eight public beta releases (version 0.13.0)"
        },
        {
          "label": "Platforms",
          "value": "Linux, macOS and Windows"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A non-developer built a cross-platform SFTP client of about 22,000 lines of Rust with 149 automated tests",
          "metric": "~22,000 lines; 149 tests",
          "quote": "ShuttleSCP is now about 22,000 lines of Rust with 149 automated tests."
        }
      ],
      "techniques": [
        "features described in plain words",
        "every feature on its own branch",
        "nothing merges until tried by hand",
        "small releases"
      ],
      "caveats": "Self-reported by the builder, who says he didn't read every line of Rust. No timeline, cost or user numbers given. Still in public beta; the Windows version is new.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "yassir-halaoui-kempt-signed-macos-app-four-days-claude-code-154-commits",
      "title": "Solo developer ships Kempt, a signed and notarised macOS app with Stripe licensing, four days after the first commit, built mostly by Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/yassir-halaoui-kempt-signed-macos-app-four-days-claude-code-154-commits/",
      "source_url": "https://yassirhalaoui.com/notes/ai-engineering/mac-app-four-days-ai-coding-agents",
      "source_name": "yassirhalaoui.com (Yassir Halaoui)",
      "source_type": "builder",
      "date": "2026-10-21",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "First commit 13 Oct to 1.0 on 17 Oct 2026 (five calendar days)"
        },
        {
          "label": "Commits",
          "value": "154 (73 on day two)"
        },
        {
          "label": "Size",
          "value": "About 20,000 lines of Swift with 122 tests"
        },
        {
          "label": "Measured speed-ups",
          "value": "Clean-up scan 8.2 s to 0.86 s; Finder tags 600 ms to 0.1 ms per file"
        }
      ],
      "duration_days": 5,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A signed, notarised Mac app with updates and licence sales shipped five calendar days after the first commit",
          "metric": "154 commits in 5 days",
          "quote": "That's 154 commits over five calendar days."
        },
        {
          "claim": "Measured speed-ups: the clean-up scan went from 8.2 s to 0.86 s, and Finder tagging from 600 ms to 0.1 ms per file",
          "metric": "8.2 s → 0.86 s"
        }
      ],
      "techniques": [
        "codex as independent reviewer",
        "eval harness on day one",
        "hand-verified labels",
        "agents run wide in parallel"
      ],
      "caveats": "Self-reported by a consultant whose post ends with a contact address. No cost figures. No sales or user numbers yet.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "rodion-keepard-family-finance-app-codex-claude-code-nine-days-643-commits",
      "title": "A CTO built a self-hosted family finance app in nine days with Codex and Claude Code, and has used it daily for two months",
      "url": "https://www.shippedbyagents.dev/stories/rodion-keepard-family-finance-app-codex-claude-code-nine-days-643-commits/",
      "source_url": "https://dev.to/rodik/i-built-a-self-hosted-family-finance-app-with-claude-code-in-9-days-heres-what-it-cost-and-what-i-2nl2",
      "source_name": "DEV Community (Rodion)",
      "source_type": "builder",
      "date": "2026-10-21",
      "tools": [
        "OpenAI Codex",
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Build time",
          "value": "Nine calendar days, about 3 to 4 hours of attention a day"
        },
        {
          "label": "Output",
          "value": "643 commits, about 200 merged PRs, 61 database migrations, 659 tests"
        },
        {
          "label": "Token cost at API prices",
          "value": "About $2,700, 72% of it re-reading context and 13% generated code; paid in practice via a $100/month subscription"
        },
        {
          "label": "Use since launch",
          "value": "Used daily by him and his wife for almost two months"
        }
      ],
      "duration_days": 9,
      "cost_usd": null,
      "prs_merged": 200,
      "claimed_benefits": [
        {
          "claim": "A self-hosted family finance app built in nine calendar days at 3 to 4 hours of attention a day, with 643 commits and 659 tests",
          "metric": "9 days; 643 commits; 659 tests",
          "quote": "In that time the repository gained 643 commits, around 200 merged pull requests, 61 database migrations and 659 tests."
        },
        {
          "claim": "About $2,700 at API prices, but $100 a month in practice on a subscription",
          "metric": "$100/month subscription"
        }
      ],
      "techniques": [
        "short dictated prompts",
        "one page of design rules",
        "component index as vocabulary",
        "screenshot script for phone and desktop",
        "parallel git worktrees"
      ],
      "caveats": "Self-reported by the builder, who says the numbers come from logs and git. Two users. He runs an agency and mentions a starter kit that grew out of the project. Code is open source at github.com/sars/keepard.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "claire-alvis-repo-fixer-scheduled-ci-pipeline-llm-fixers-150-migration-prs-overnight",
      "title": "Repo fixer: a scheduled CircleCI pipeline with LLM fixers opens 150 migration PRs across an org overnight and keeps fixing new repos",
      "url": "https://www.shippedbyagents.dev/stories/claire-alvis-repo-fixer-scheduled-ci-pipeline-llm-fixers-150-migration-prs-overnight/",
      "source_url": "https://www.confidentcommit.com/p/automate-code-migrations-llm-pipeline",
      "source_name": "Confident Commit (Claire Alvis)",
      "source_type": "builder",
      "date": "2026-10-21",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Migration PRs",
          "value": "150 open PRs across the org by the next day"
        },
        {
          "label": "Previous approach",
          "value": "Over a hundred PRs opened by hand, one repo at a time"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "One prompt produced 150 open migration PRs across the org by the next day",
          "metric": "150 PRs overnight",
          "quote": "by the next day there were 150 open PRs across the org with the change already made"
        },
        {
          "claim": "Engineers now ask to add their own fixers without being prompted"
        }
      ],
      "techniques": [
        "cheap detection pass before llm",
        "scheduled fixers run forever",
        "draft prs from a bot user",
        "bulk human review before release",
        "sme writes fixer prompt once"
      ],
      "caveats": "Self-reported. The post does not name the employer, the LLM or coding agent used for fixes, or merge rates for the 150 PRs. The Claude skills are for managing the system, not necessarily for the fixes.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stevan-bogosavljevic-lokalbot-150000-line-swift-app-ported-to-rust-three-days-agents",
      "title": "Solo developer ports LokalBot, a 150,000-line Swift Mac app, to Rust for Windows and Linux in about 53 hours with Claude Code agents",
      "url": "https://www.shippedbyagents.dev/stories/stevan-bogosavljevic-lokalbot-150000-line-swift-app-ported-to-rust-three-days-agents/",
      "source_url": "https://stevan.bogosavljevic.me/blog/posts/porting-lokalbot-to-linux-and-windows/",
      "source_name": "Field Notes (Stevan Bogosavljević)",
      "source_type": "builder",
      "date": "2026-10-20",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Source app size",
          "value": "About 150,000 lines of Swift"
        },
        {
          "label": "Time",
          "value": "About 53 hours from first commit to the 0.11.0 release"
        },
        {
          "label": "Human input",
          "value": "About 35 messages and about three hours at the Mac"
        },
        {
          "label": "CI time",
          "value": "Full run from about 32 to 17 minutes on self-hosted runners, then about 9 after a Windows fix"
        },
        {
          "label": "Cost",
          "value": "No API bill; ran on a regular subscription (nearly all weekly usage two weeks running)"
        }
      ],
      "duration_days": 2.208,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Windows and Linux versions released about 53 hours after the first commit, with about 35 messages and three hours at the Mac from the author",
          "metric": "~53 hours",
          "quote": "Version 0.11.0 went out at 12:22 on Tuesday, about 53 hours after the first commit."
        },
        {
          "claim": "A full CI run cut from about 32 minutes to 17",
          "metric": "32 → 17 min",
          "quote": "A full CI run went from about 32 minutes to 17"
        }
      ],
      "techniques": [
        "read-only agents write specs of the existing app first",
        "lead agent writes the brief",
        "six implementation agents in parallel",
        "ci screenshots every page"
      ],
      "caveats": "Self-reported. The author didn't review the code himself and judged it from screenshots, demo videos and CI. Windows and Linux builds are alpha, unsigned, and some Mac features aren't ported yet. No line count for the Rust port.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "uber-autoclean-agentic-feature-flag-cleanup-16000-flags-1-million-lines",
      "title": "Uber's AutoClean agents remove stale feature flags across its monorepos, with over 16,000 flags and 1,000,000 lines cleaned",
      "url": "https://www.shippedbyagents.dev/stories/uber-autoclean-agentic-feature-flag-cleanup-16000-flags-1-million-lines/",
      "source_url": "https://www.uber.com/us/en/blog/ai-code-cleanup/",
      "source_name": "Uber Engineering blog (Daniel Ramos, Diego Marcilio)",
      "source_type": "builder",
      "date": "2026-10-20",
      "tools": [],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Flags cleaned",
          "value": "\"more than 16,000 flags\" (\"Over the past few years\")"
        },
        {
          "label": "Code removed",
          "value": "\"over 1,000,000 lines of code\""
        },
        {
          "label": "Scope",
          "value": "\"thousands of feature flags across our mobile apps and back-end services\""
        },
        {
          "label": "Human gate",
          "value": "Every diff goes to the owning team for review; \"a human reviewer makes the final decision\""
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 16,000 stale feature flags and over 1,000,000 lines of code cleaned up over the past few years",
          "metric": "16,000+ flags; 1M+ lines",
          "quote": "Over the past few years, AutoClean has cleaned up more than 16,000 flags and cleaned over 1,000,000 lines of code."
        }
      ],
      "techniques": [
        "agent calls a deterministic engine over mcp",
        "clean, tested diffs ready for human review",
        "runs continuously across monorepos",
        "owning team makes the final decision"
      ],
      "caveats": "Uber's own engineering blog about its in-house system, so self-reported, though Uber doesn't sell the tool. The 16,000 flags and 1,000,000 lines are totals \"over the past few years\" and the post doesn't say how many came from the agentic version versus earlier PolyglotPiranha-based automation. It names no AI model or vendor, and gives no cost, time per cleanup, acceptance or abandonment rates, or count of diffs rejected by reviewers.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "kassebaum-engineering-agents-review-and-merge-136-prs-ten-days-risk-tiers",
      "title": "Kassebaum Engineering lets agents review and merge routine code: 136 PRs across four repos in ten days, with people merging the riskiest tier",
      "url": "https://www.shippedbyagents.dev/stories/kassebaum-engineering-agents-review-and-merge-136-prs-ten-days-risk-tiers/",
      "source_url": "https://kassebaumengineering.com/insights/agent-orchestration-trusted-pair-programming/",
      "source_name": "Kassebaum Engineering insights (William R. Kassebaum, PE)",
      "source_type": "builder",
      "date": "2026-10-20",
      "tools": [
        "Claude Code",
        "Cursor",
        "Grok CLI"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Merged PRs",
          "value": "136 across four repositories over ten days in October 2026"
        },
        {
          "label": "Review coverage",
          "value": "Nearly every PR carries a written review verdict"
        },
        {
          "label": "Risk tiers",
          "value": "Four (T0 docs to T3 money/auth); T3 needs two reviewers and a human merge"
        }
      ],
      "duration_days": 10,
      "cost_usd": null,
      "prs_merged": 136,
      "claimed_benefits": [
        {
          "claim": "136 agent-written pull requests reviewed and merged across four repos in ten days, nearly all with a written review verdict",
          "metric": "136 PRs in 10 days",
          "quote": "Over ten days in October 2026, 136 pull requests went through this process and merged across four repositories."
        }
      ],
      "techniques": [
        "a different agent reviews each pr",
        "machine-readable review verdict",
        "merge gate",
        "risk tiers by file",
        "human merges money and auth changes"
      ],
      "caveats": "Small firm writing about its own setup, and the post ends with a sales pitch. The author says they have no time-saved figure. No security scanning in CI yet, and the cross-model review panel is built but not live. Uses Grok-based agents, including a Grok Bot engineering agent.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cisco-pubhub-go-rewrite-claude-copilot-cursor-nine-months-in-five-and-a-half",
      "title": "Cisco DevNet's 12-engineer team ships a mandatory PubHub rewrite in Go in about five and a half months instead of nine, using Claude, Copilot and Cursor",
      "url": "https://www.shippedbyagents.dev/stories/cisco-pubhub-go-rewrite-claude-copilot-cursor-nine-months-in-five-and-a-half/",
      "source_url": "https://blogs.cisco.com/cisco-devnet/ai-assisted-development-rebuilding-pubhub-part-1",
      "source_name": "Cisco Blogs, DevNet (Koushik Dhar)",
      "source_type": "builder",
      "date": "2026-10-19",
      "tools": [
        "Claude",
        "GitHub Copilot",
        "Cursor"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Schedule",
          "value": "Nine planned months shipped in about five and a half (39% reduction against the pre-AI estimate)"
        },
        {
          "label": "Effort",
          "value": "108 planned engineer-months of scope shipped in roughly 66, with 12 engineers"
        },
        {
          "label": "Throughput",
          "value": "1.64x whole-project average; higher after Claude arrived in June"
        },
        {
          "label": "Quality",
          "value": "No major issues since reaching production"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Nine planned months shipped in about five and a half, a 39% schedule cut against the team's pre-AI estimate",
          "metric": "-39% schedule",
          "quote": "Nine months of planned work shipped in about five and a half: a 39% schedule reduction against the team’s own pre-AI estimate."
        },
        {
          "claim": "12 engineers producing the output of about 20, a 1.64x throughput gain averaged over the project",
          "metric": "1.64x",
          "quote": "12 engineers producing the output of about 20"
        },
        {
          "claim": "Reached production roughly three and a half months ahead of the original December plan"
        }
      ],
      "techniques": [
        "architecture and component reference generated from the old app",
        "proof of concept first",
        "hand-conversion pilot dropped as too slow"
      ],
      "caveats": "Written by the team itself. Compared against their own pre-AI estimate, not a measured baseline. Claude only joined in June, so the first two months ran at the old pace. No cost, PR or line counts. Engineering detail is promised in a separate AI for SDLC series.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "loft-three-person-team-claude-code-spec-first-53-hour-release",
      "title": "Three people build Loft with coding agents over three months, then ship its open format 53 hours after the first commit",
      "url": "https://www.shippedbyagents.dev/stories/loft-three-person-team-claude-code-spec-first-53-hour-release/",
      "source_url": "https://askloft.ai/insights/articles/building-loft-with-claude-code",
      "source_name": "Loft (askloft.ai)",
      "source_type": "builder",
      "date": "2026-10-16",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Team",
          "value": "3 people"
        },
        {
          "label": "V1 duration",
          "value": "~3 months (July 10 – Oct 16)"
        },
        {
          "label": "V1 commits",
          "value": "Nearly 2,000 (merges excluded)"
        },
        {
          "label": "V2 public release",
          "value": "53 hours after first commit"
        },
        {
          "label": "V2 commits",
          "value": "~350 in six days"
        },
        {
          "label": "Spec size",
          "value": "~5,000 nodes (~1,000 requirements, 3,500 acceptance criteria)"
        },
        {
          "label": "Agent share",
          "value": "90%+ of V1's October commits co-authored by Claude"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "The open format went public 53 hours after its first commit",
          "metric": "53 hours",
          "quote": "The open Loft format and its loft command line went public 53 hours after their first commit."
        },
        {
          "claim": "The everyday test gate went from over 40 minutes to under 10",
          "metric": "40+ → <10 min",
          "quote": "The same loop cut the everyday test gate from over 40 minutes to under 10."
        }
      ],
      "techniques": [
        "spec first",
        "deterministic gates",
        "fresh reviewing agents",
        "function-length and complexity limits",
        "people ratify the spec"
      ],
      "caveats": "Written by the company about its own product, which it sells as a way to run agents. No cost, token or code-size figures. The fast V2 release rested on three months of earlier V1 work.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stack-overflow-for-agents-two-engineers-codex-few-weeks-connectors-4-days",
      "title": "Two Stack Overflow engineers build Stack Overflow for Agents in a few weeks with Codex, and connectors drop from 3–6 weeks to 4 days",
      "url": "https://www.shippedbyagents.dev/stories/stack-overflow-for-agents-two-engineers-codex-few-weeks-connectors-4-days/",
      "source_url": "https://openai.com/index/stack-overflow-chatgpt-plugin/",
      "source_name": "OpenAI customer story",
      "source_type": "vendor",
      "date": "2026-10-16",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Team and time",
          "value": "Two engineers, \"a few weeks\" to a full product"
        },
        {
          "label": "Connector build time",
          "value": "3–6 weeks on average, 4 days with Codex, most recently 1 day"
        },
        {
          "label": "Enterprise path to market",
          "value": "Aiming for three months rather than a year"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Two engineers built a full new product in a few weeks",
          "quote": "Within a few weeks, Alex says, they had a full-fledged product."
        },
        {
          "claim": "A connector that takes 3 to 6 weeks on average was built in 4 days, and more recently in one day",
          "metric": "3–6 weeks → 4 days",
          "quote": "On average, it takes us 3 to 6 weeks to build one connector. With Codex, we were able to build it within 4 days."
        }
      ],
      "techniques": [
        "codex given a development environment and a plan",
        "two engineers",
        "chatgpt plugin for distribution"
      ],
      "caveats": "OpenAI customer story built on quotes from one Stack Overflow executive. 'A few weeks' isn't exact, and there are no code, PR or cost numbers. The signup and agent-creation figures in the headline are about the ChatGPT plugin, not the build.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "navan-spring-boot-4-migration-8-apps-15-libraries-one-week-cursor-cloud-agents",
      "title": "Navan moves 8 apps and 15 libraries to Spring Boot 4 in about a week with Cursor cloud agents, against over a month for a comparable migration two years earlier",
      "url": "https://www.shippedbyagents.dev/stories/navan-spring-boot-4-migration-8-apps-15-libraries-one-week-cursor-cloud-agents/",
      "source_url": "https://cursor.com/blog/navan",
      "source_name": "Cursor customer story",
      "source_type": "vendor",
      "date": "2026-10-16",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Migration time",
          "value": "About a week for 8 apps and 15 libraries (previous comparable migration: over a month)"
        },
        {
          "label": "Cloud agent users",
          "value": "Roughly 200 people, a 10x increase since April 2026"
        },
        {
          "label": "Share of merged PRs",
          "value": "36% of merged core-repo PRs authored with Cursor over 30 days"
        }
      ],
      "duration_days": 7,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A Spring Boot 4 migration of 8 apps and 15 libraries finished in about a week without freezing development; a comparable one two years earlier took over a month",
          "metric": "~1 week vs over a month",
          "quote": "An 8-app migration finishes in a week instead of a month"
        },
        {
          "claim": "36% of merged core-repo pull requests authored with Cursor over a 30-day period",
          "metric": "36%"
        }
      ],
      "techniques": [
        "plan the upgrade once",
        "cloud agents on self-hosted machines",
        "one component per agent in parallel",
        "agents use internal tooling and browsers"
      ],
      "caveats": "Vendor customer story with a sales pitch at the end. The month-long baseline is one earlier migration, not a controlled comparison. No figures on review effort, defects or rollbacks.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "claude-code-team-code-review-merged-prs-doubled-reverts-down-40-percent",
      "title": "Claude Code team doubles merged pull requests from June to September while the revert rate falls by more than 40%, by moving most review to agents",
      "url": "https://www.shippedbyagents.dev/stories/claude-code-team-code-review-merged-prs-doubled-reverts-down-40-percent/",
      "source_url": "https://claude.dev/blog/solving-code-review-at-anthropic/",
      "source_name": "claude.dev blog (Addy Osmani, Morgan Westlee Lunt, Qing Wang)",
      "source_type": "vendor",
      "date": "2026-10-15",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Merged PRs",
          "value": "More than doubled from June to September"
        },
        {
          "label": "Revert rate",
          "value": "Share of merged PRs reverted fell by more than 40%"
        },
        {
          "label": "Substantive review comments",
          "value": "Share of internal PRs getting them went from 16% to 54% at the research-preview launch"
        },
        {
          "label": "Agent-written code",
          "value": "About 90% of Claude Code's own code"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Merged PRs more than doubled from June to September while the share reverted fell by more than 40%",
          "metric": "2x+ PRs; -40% reverts",
          "quote": "From June to September, merged pull requests to Claude Code's codebase more than doubled while the share that had to be reverted fell by more than 40 percent."
        },
        {
          "claim": "Share of internal PRs getting substantive review comments went from 16% to 54%",
          "metric": "16% → 54%",
          "quote": "the share of internal PRs getting substantive review comments went from 16% to 54%"
        }
      ],
      "techniques": [
        "claude verifies its own work",
        "fresh-context subagent reviews the diff",
        "ai review findings cross-checked by other agents",
        "blocking security review",
        "review built around human time"
      ],
      "caveats": "Vendor post about its own product, which also promotes Code Review for Claude Code. No absolute PR counts, incident numbers or cost. About review and merging rather than one build.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "meta-forum-groups-app-human-on-the-loop-agents-feature-parity-eight-weeks",
      "title": "Meta's small Forum team builds a new Groups app from scratch with agents running the build-test loop, reaching feature parity in eight weeks",
      "url": "https://www.shippedbyagents.dev/stories/meta-forum-groups-app-human-on-the-loop-agents-feature-parity-eight-weeks/",
      "source_url": "https://engineering.fb.com/2026/10/15/ml-applications/how-meta-built-forum-ai-native-engineering/",
      "source_name": "Engineering at Meta (Shengbo Guo, Bohou Luo and others)",
      "source_type": "builder",
      "date": "2026-10-15",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time to feature parity",
          "value": "Core feature parity with Facebook Groups in eight weeks"
        },
        {
          "label": "Memory",
          "value": "Agent stress-testing loop found and fixed a memory issue during development, roughly halving memory use and page faults"
        }
      ],
      "duration_days": 56,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Core feature parity with Facebook Groups in eight weeks",
          "metric": "8 weeks",
          "quote": "Forum reached core feature parity with Facebook Groups in eight weeks"
        },
        {
          "claim": "An agent stress-testing loop caught a memory issue during development, roughly halving memory use and page faults",
          "metric": "~50% less memory",
          "quote": "roughly halving memory use and page faults"
        }
      ],
      "techniques": [
        "human-on-the-loop",
        "agent plans, implements, builds and tests",
        "engineer sets the goal and signs off",
        "agent stress-testing runs",
        "continuous verification"
      ],
      "caveats": "Meta's own engineering blog. It names no agent, model or vendor, and gives no team size, cost, PR counts or code volume. 'Feature parity in eight weeks' has no pre-agent baseline to compare against.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "foxbox-claude-code-production-marketing-site-14-days-76-commits",
      "title": "Foxbox Digital ships a production marketing site with two engineers in 14 days using Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/foxbox-claude-code-production-marketing-site-14-days-76-commits/",
      "source_url": "https://foxbox.com/blog/14-days-76-commits-one-production-website-what-happens-when-you-build-with-claude-code",
      "source_name": "Foxbox Digital blog",
      "source_type": "builder",
      "date": "2026-10-14",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Calendar time",
          "value": "2 weeks"
        },
        {
          "label": "Team",
          "value": "2 engineers"
        },
        {
          "label": "Lines of code",
          "value": "~9,200"
        },
        {
          "label": "Source files",
          "value": "72 (TS/TSX/JS)"
        },
        {
          "label": "Commits / PRs merged",
          "value": "76 / 76"
        },
        {
          "label": "Pages",
          "value": "12"
        },
        {
          "label": "CMS schemas",
          "value": "11"
        }
      ],
      "duration_days": 14,
      "cost_usd": null,
      "prs_merged": 76,
      "claimed_benefits": [
        {
          "claim": "Two engineers shipped a production marketing site of about 9,200 lines in two weeks, with 76 commits and 76 PRs merged",
          "metric": "2 weeks; 76 PRs",
          "quote": "Here’s a concrete snapshot of what two engineers shipped in two weeks"
        },
        {
          "claim": "A component scaffolded, reviewed and done in under an hour",
          "quote": "we were done in under an hour"
        }
      ],
      "techniques": [
        "every commit through a reviewed pr",
        "describe, build, review loop"
      ],
      "caveats": "Written by the agency that did the work, partly as marketing for its services. No cost, token usage or pre-AI estimate is given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "sparq-22-agent-platform-claude-code-34000-lines",
      "title": "Sparq builds a 22-agent internal platform with Claude Code: 34,000 lines across 453 commits",
      "url": "https://www.shippedbyagents.dev/stories/sparq-22-agent-platform-claude-code-34000-lines/",
      "source_url": "https://www.teamsparq.com/insights/building-agentic-platform-with-claude-code/",
      "source_name": "Sparq Insights (build log)",
      "source_type": "builder",
      "date": "2026-10-14",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Lines of code",
          "value": "~34,000 Python"
        },
        {
          "label": "Commits / PRs",
          "value": "453 / 159"
        },
        {
          "label": "Repositories",
          "value": "4"
        },
        {
          "label": "Specialist agents",
          "value": "22"
        },
        {
          "label": "Duration",
          "value": "May 2026 to Oct 2026"
        },
        {
          "label": "Baseline prep time",
          "value": "3–4 hours per client interaction"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "specialist agents behind an orchestrator",
        "scheduled autonomous jobs",
        "portable skills",
        "bugfix checkpoint log"
      ],
      "caveats": "Sparq is a consultancy and the post doubles as marketing for its AI readiness services. The authors say the numbers measure delivery volume from GitHub history, not business impact, and that there was no comparison build without Claude Code. Team size isn't stated.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "anthropic-scion-in-house-version-control-seven-engineers-five-months-claude",
      "title": "Seven Anthropic engineers use Claude to build Scion, an in-house version control system, and move the monorepo off git in five months",
      "url": "https://www.shippedbyagents.dev/stories/anthropic-scion-in-house-version-control-seven-engineers-five-months-claude/",
      "source_url": "https://www.anthropic.com/engineering/redesigning-source-control",
      "source_name": "Anthropic Engineering blog",
      "source_type": "vendor",
      "date": "2026-10-13",
      "tools": [
        "Claude"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Team",
          "value": "7 people"
        },
        {
          "label": "Time",
          "value": "5 months to build and fully migrate"
        },
        {
          "label": "Merge volume",
          "value": "About 2,500 merges a day in May 2026 to over 30,000 a day in September"
        },
        {
          "label": "Infrastructure",
          "value": "Footprint \"200x smaller\" than the old git mirroring setup; CI got a 25% CPU capacity win"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "New version control infrastructure 200x smaller than the old git mirroring setup in CPU and memory",
          "metric": "200x",
          "quote": "Our Scion infrastructure footprint is now 200x smaller than our previous git mirroring infrastructure in terms of CPU and memory requirements."
        },
        {
          "claim": "CI and research jobs check out code in seconds rather than minutes, and CI got a 25% CPU capacity win",
          "metric": "25% CPU",
          "quote": "CI alone realized a 25% CPU capacity win"
        },
        {
          "claim": "One rewrite by Claude used 20x less memory and ran 4x faster",
          "metric": "20x less memory; 4x faster",
          "quote": "it produced a version that consumed 20x less memory and was 4x faster"
        }
      ],
      "techniques": [
        "learning from piper and sapling",
        "claude writes most of the code",
        "migration while merge volume grew 10x"
      ],
      "caveats": "Anthropic writing about its own model, so a vendor source. The post doesn't say how much of the code Claude wrote, give a line count or token cost, and notes Anthropic spends far more on tokens internally than other companies would. The infrastructure numbers describe Scion's performance, not the agent's work.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "google-gemini-4-argon-overnight-adk-python-optimization-40-percent",
      "title": "Gemini 4 Argon runs unattended for 11 hours on the 205,000-line ADK Python codebase; 2 of its 6 optimizations ship",
      "url": "https://www.shippedbyagents.dev/stories/google-gemini-4-argon-overnight-adk-python-optimization-40-percent/",
      "source_url": "https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/",
      "source_name": "Google Developers Blog",
      "source_type": "vendor",
      "date": "2026-10-13",
      "tools": [
        "Google Antigravity"
      ],
      "models": [
        "Gemini 4 Argon"
      ],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "~205,000 lines, ~800 source files, ~17,000 unit tests"
        },
        {
          "label": "Run length",
          "value": "11 hours unattended (after a 2-hour first round)"
        },
        {
          "label": "Proposals",
          "value": "6 optimizations across 32 files; 2 shipped, 4 rejected in review"
        },
        {
          "label": "Session prep overhead",
          "value": "About 40% lower in long sessions"
        },
        {
          "label": "Package import time",
          "value": "21% faster"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Session prep time cut by about 40% in long sessions and cold package imports 21% faster, after an 11-hour unattended run",
          "metric": "-40% prep; 21% faster imports",
          "quote": "cut ADK's in-memory session prep time by about 40% in long sessions and sped up cold package imports by 21%"
        },
        {
          "claim": "A 158-turn run went from 65.6 to 56.1 seconds",
          "metric": "65.6 s → 56.1 s",
          "quote": "cuts the full 158-turn run from 65.6 seconds to 56.1 seconds"
        }
      ],
      "techniques": [
        "one-paragraph goal",
        ".agents/skills rulefiles",
        "/goal mode overnight run",
        "humans cut 4 of 6 proposals in review",
        "holdout benchmarks"
      ],
      "caveats": "Vendor launch post for its own model. Time to first model request in a new process stayed flat (3.07 s to 3.09 s). No cost figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "lamplit-labs-dotnet-framework-48-to-dotnet-8-migration-coding-agent-11-days",
      "title": "Two engineers and a coding agent move a 41,000-line .NET Framework 4.8 service to .NET 8 in 11 working days, against a 7 to 9 engineer-week estimate",
      "url": "https://www.shippedbyagents.dev/stories/lamplit-labs-dotnet-framework-48-to-dotnet-8-migration-coding-agent-11-days/",
      "source_url": "https://blogs.lamplitlabs.com/posts/ai-sdlc-dotnet-framework-to-dotnet8-migration-coding-agent/",
      "source_name": "Lamplit Labs blog (Manish Tiwari)",
      "source_type": "builder",
      "date": "2026-10-12",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "41k-line .NET Framework 4.8 service"
        },
        {
          "label": "Time",
          "value": "11 working days, two engineers part-time, vs a 7 to 9 engineer-week estimate"
        },
        {
          "label": "Agent PRs",
          "value": "9 opened, 7 merged (4 with substantial human edits), 2 closed"
        },
        {
          "label": "Agent share",
          "value": "71% of changed lines in the final diff"
        },
        {
          "label": "EF queries",
          "value": "121 converted by the agent, 31 by hand"
        },
        {
          "label": "Production defects, first month",
          "value": "2, both in hand-written code"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 7,
      "claimed_benefits": [
        {
          "claim": "Migration done in 11 working days against a 7 to 9 engineer-week estimate",
          "metric": "11 working days vs 7-9 engineer-weeks"
        },
        {
          "claim": "Agent wrote 71% of changed lines in the final diff",
          "metric": "71%"
        },
        {
          "claim": "No first-month production defects in agent-written code that passed the gates",
          "quote": "None in agent-written code that passed the gates"
        }
      ],
      "techniques": [
        "scope contract per pr",
        "merge gates with mutation score",
        "test baseline before migration",
        "upgrade assistant report as agent input",
        "humans keep design decisions"
      ],
      "caveats": "Self-reported by the team. The post does not name the coding agent or model. No cost figures. The author notes that the clean record of agent code says more about the gates and ported tests than about the agent.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "smarter-django-platform-claude-code-19-releases-18-days",
      "title": "Claude Code ships 19 minor releases of the Smarter Django platform in 18 days, adding about 238,000 lines",
      "url": "https://www.shippedbyagents.dev/stories/smarter-django-platform-claude-code-19-releases-18-days/",
      "source_url": "https://blog.smarter.sh/2026/10/10/claude-code-large-django-codebase/",
      "source_name": "The Smarter Project blog",
      "source_type": "builder",
      "date": "2026-10-10",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Releases",
          "value": "19 minor and 45 patch releases in 18 days"
        },
        {
          "label": "Road map",
          "value": "Finished in 8 days over about 20 prompts"
        },
        {
          "label": "Lines added by Claude Code",
          "value": "About 238,000 in 18 days (author: about 43,000, mostly docs)"
        },
        {
          "label": "Codebase size",
          "value": "About 275,000 lines of Python excluding migrations (17 Oct)"
        },
        {
          "label": "Share of production code by Claude Code",
          "value": "About a quarter on 10 Oct, about a third on 17 Oct"
        }
      ],
      "duration_days": 18,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "The rest of the 2026 road map finished in 8 days over about 20 prompts, then 19 minor releases in 18 days",
          "metric": "8 days; 19 releases in 18 days",
          "quote": "It finished in 8 days, over about 20 prompts"
        },
        {
          "claim": "About 25 times the author's long-run pace of writing code",
          "metric": "~25x",
          "quote": "about 25 times the author’s long-run pace"
        },
        {
          "claim": "Roughly 130 bugs fixed in the first week and overall test coverage raised to 96%",
          "metric": "96% coverage"
        }
      ],
      "techniques": [
        "high test coverage before handing over",
        "curated docstrings as context",
        "pre-commit hooks as guard rails",
        "written rules in the environment"
      ],
      "caveats": "Self-reported by the project author, who notes lines of code are a poor productivity measure. Line counts come from GitHub contributor stats. The post covers data up to 17 October.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cockroach-labs-molt-sinai-agent-hospital-db2-support-two-days-4172-dollars",
      "title": "Cockroach Labs' agent pipeline adds IBM Db2 support to its migration tool in under two days for $4,172 in tokens, then lands 1,238 PRs in five months",
      "url": "https://www.shippedbyagents.dev/stories/cockroach-labs-molt-sinai-agent-hospital-db2-support-two-days-4172-dollars/",
      "source_url": "https://www.cockroachlabs.com/blog/experiment-running-hospital-code/",
      "source_name": "Cockroach Labs blog (Adam Storm, Rafi Shamim)",
      "source_type": "builder",
      "date": "2026-10-09",
      "tools": [
        "Claude"
      ],
      "models": [
        "Claude Opus",
        "Claude Sonnet",
        "Claude Fable"
      ],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Db2 support time",
          "value": "\"less than two days\" vs 9 months for Oracle support in 2024"
        },
        {
          "label": "Db2 token cost",
          "value": "$4,172 vs \"around $160K in engineering time\" for Oracle"
        },
        {
          "label": "Db2 work",
          "value": "32 sub-issues, 27 PRs merged, 55 send-backs by reviewers, 9 escalations (2 to a human)"
        },
        {
          "label": "Five-month output",
          "value": "\"well over a million lines of code\", 1,238 PRs, 7 reverts (21 Apr to 11 Sep)"
        },
        {
          "label": "Total token spend",
          "value": "\"just over $135,000\" in Claude tokens, about $84 per issue"
        },
        {
          "label": "Self-built",
          "value": "About 1,340 of the Sinai repo's 1,570 commits (85%) authored by the pipeline"
        }
      ],
      "duration_days": null,
      "cost_usd": 4172,
      "prs_merged": 27,
      "claimed_benefits": [
        {
          "claim": "Db2 support took less than two days and $4,172 in tokens, where Oracle support in 2024 took 9 months and about $160K in engineering time: 164x faster and 38x cheaper",
          "metric": "164x faster; 38x cheaper",
          "quote": "The token bill was $4172: 164x faster, and 38x cheaper."
        },
        {
          "claim": "Well over a million lines landed in just under five months, in changes small enough for agents to review",
          "quote": "In just under 5 months, the hospital pipeline landed well over a million lines of code"
        },
        {
          "claim": "A full day of meetings can lead to 10 merged PRs",
          "quote": "A full day of meetings can now lead to 10 high quality merged PRs."
        }
      ],
      "techniques": [
        "teaching hospital model",
        "reproduce the problem before any code",
        "separate agent approves the plan",
        "decompose changes over 1,000 lines",
        "safety review after three rework rounds"
      ],
      "caveats": "Self-reported on Cockroach Labs' own blog. The Db2 sprint ran in a mirror repo with no human review, and the authors say they are still confirming the Db2 support is correct. The Oracle comparison is against a 2024 human project of different scope, so the 164x and 38x figures are rough. Human approval was turned on for every merge after the first week.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dennis-schoneberg-scumble-desktop-image-editor-claude-code-five-weeks",
      "title": "Solo developer ships Scumble, a desktop image editor in the Microsoft Store, five weeks after the first commit with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/dennis-schoneberg-scumble-desktop-image-editor-claude-code-five-weeks/",
      "source_url": "https://dev.to/denrakeiw/i-built-a-desktop-image-editor-in-five-weeks-with-claude-code-46g2",
      "source_name": "DEV Community (Dennis Schöneberg)",
      "source_type": "builder",
      "date": "2026-10-09",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Five weeks from first commit to release 43"
        },
        {
          "label": "Releases",
          "value": "43 (0.1.0 on 9 Sep to 0.1.44 on 9 Oct)"
        },
        {
          "label": "Commits",
          "value": "632 in the app, 137 in the ComfyUI node"
        },
        {
          "label": "App code",
          "value": "~83,000 lines of JavaScript, ~2,900 lines of Rust"
        },
        {
          "label": "Test and tooling code",
          "value": "~83,000 lines"
        }
      ],
      "duration_days": 35,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A desktop image editor on its 43rd release and in the Microsoft Store five weeks after the first commit",
          "metric": "5 weeks; 43 releases",
          "quote": "Today, five weeks after that first commit, Scumble is on its 43rd release, in the Microsoft Store, and edits 15,000-pixel pictures without freezing."
        }
      ],
      "techniques": [
        "one section per session, then /clear",
        "decisions list the agent may not reopen",
        "pruned project notes file",
        "heavier tests only for code that can lose data"
      ],
      "caveats": "Self-reported by the developer. No cost figures. Line counts are approximate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "unsw-27000-agent-prs-deployment-infrastructure-changes-merge-unreviewed",
      "title": "Study of 26,999 PRs: agent changes that touch both app code and deployment config wait 3.24x longer when reviewed, yet 43–56% merge with no human review",
      "url": "https://www.shippedbyagents.dev/stories/unsw-27000-agent-prs-deployment-infrastructure-changes-merge-unreviewed/",
      "source_url": "https://arxiv.org/abs/2610.12849",
      "source_name": "arXiv 2610.12849",
      "source_type": "independent",
      "date": "2026-10-09",
      "tools": [
        "Claude Code",
        "OpenAI Codex",
        "GitHub Copilot",
        "Cursor",
        "Devin",
        "Jules"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs studied",
          "value": "26,999 (AIDev v5: 21,855; MOSAIC-3M: 5,144)"
        },
        {
          "label": "Dual-scope share",
          "value": "4.48% to 4.55% of agent PRs"
        },
        {
          "label": "Review delay",
          "value": "Median +13.13 hours, 3.24x time-to-merge for reviewed dual-scope PRs"
        },
        {
          "label": "Unreviewed merges",
          "value": "43.1% to 56.2% of dual-scope PRs merge without inspection, median under 31 minutes"
        },
        {
          "label": "Review odds",
          "value": "Agent identity lowers odds of human review by 63.7% vs. human peers"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "dual-scope prs",
        "nearest-neighbor matching",
        "aidev v5 and mosaic-3m datasets"
      ],
      "caveats": "Preprint, not yet peer reviewed. One author is listed as an independent researcher. Open-source repositories only. Measures merge and review outcomes, not whether the merged changes broke anything. The six agents are taken from the paper's text; check the per-agent breakdown before publishing.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "block-claude-fable-orchestrates-thousand-pr-migrations-two-human-approvals",
      "title": "Block uses Claude Fable to orchestrate code migrations of thousands of pull requests, with two humans approving every deploy",
      "url": "https://www.shippedbyagents.dev/stories/block-claude-fable-orchestrates-thousand-pr-migrations-two-human-approvals/",
      "source_url": "https://claude.com/resources/articles/how-block-orchestrates-claude-fable-across-thousands-of-pull-requests",
      "source_name": "Anthropic (interview with Bradley Axen, Block)",
      "source_type": "vendor",
      "date": "2026-10-08",
      "tools": [],
      "models": [
        "Claude Fable 5",
        "Claude Opus",
        "Claude Sonnet"
      ],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Migration size",
          "value": "\"hundreds or thousands of individual pull requests across multiple repos and millions of lines of code\""
        },
        {
          "label": "PRs per migration",
          "value": "\"we might merge a thousand pull requests orchestrated by Fable\""
        },
        {
          "label": "Worker models",
          "value": "\"up to dozens of smaller, more cost-effective models like Opus or Sonnet\""
        },
        {
          "label": "Human gate",
          "value": "Merges to main and production deploys stay with people; \"two humans have to approve the deploy\""
        },
        {
          "label": "Tech debt passes",
          "value": "\"daily or weekly passes\" over a codebase \"in the neighborhood of a million lines of code\""
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A single migration might merge a thousand pull requests orchestrated by Fable",
          "quote": "In any given migration, we might merge a thousand pull requests orchestrated by Fable"
        }
      ],
      "techniques": [
        "frontier model orchestrates smaller worker models",
        "human approval for merges to main and production",
        "picking the right model for the right task"
      ],
      "caveats": "This is an interview published by Anthropic, which sells Claude, not Block's own write-up. It gives no timelines, costs, token counts, defect rates or before-and-after comparisons, and names no specific migration. The cost claim (\"without a drop in quality\") is described only as early results, with no figures.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-10-08 from the page's article:published_time and JSON-LD datePublished."
    },
    {
      "slug": "lucas-firmo-ios-app-swift-6-strict-concurrency-10-days-agent-fleet",
      "title": "Lucas Firmo's agent fleet moves a production iOS app to Swift 6 strict concurrency in 10 days, with zero warnings left",
      "url": "https://www.shippedbyagents.dev/stories/lucas-firmo-ios-app-swift-6-strict-concurrency-10-days-agent-fleet/",
      "source_url": "https://www.bitswithlucas.com/blog/ai-agents-on-production-code/",
      "source_name": "Bits with Lucas (Lucas Firmo)",
      "source_type": "builder",
      "date": "2026-10-08",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "10 days for the Swift 6 strict concurrency migration"
        },
        {
          "label": "Result",
          "value": "\"zero concurrency warnings\", feature work never stopped"
        },
        {
          "label": "Tests",
          "value": "Test suite 25% larger than at the start"
        }
      ],
      "duration_days": 10,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Swift 6 strict concurrency migration done in 10 days with zero concurrency warnings, feature work never stopped, and a test suite 25% larger",
          "metric": "10 days; 0 warnings; +25% tests",
          "quote": "We moved our iOS app to Swift 6 strict concurrency in 10 days."
        }
      ],
      "techniques": [
        "git worktree per task",
        "plan and log file on disk",
        "agents claim a ticket first",
        "claude code hooks",
        "pin behaviour with a test before changing it"
      ],
      "caveats": "A short personal blog post. It doesn't name the app, team or employer, the codebase size, the number of agents, the cost, or which second model does the reviews. Claude Code is named only as the tool whose hooks enforce the rules.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "esem-2026-669-developers-agent-adoption-longer-pr-descriptions-more-testing",
      "title": "ESEM 2026 study of 669 developers: after adopting coding agents they write longer PR descriptions and test more, with no drop in effort",
      "url": "https://www.shippedbyagents.dev/stories/esem-2026-669-developers-agent-adoption-longer-pr-descriptions-more-testing/",
      "source_url": "https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.ESEM.2026.67",
      "source_name": "ESEM 2026 (LIPIcs)",
      "source_type": "independent",
      "date": "2026-10-05",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Developers studied",
          "value": "669 across 103 repositories, plus 228 controls"
        },
        {
          "label": "PR description length",
          "value": "+76.8% after agent adoption"
        },
        {
          "label": "Coding effort",
          "value": "Stable"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "PR descriptions got 76.8% longer after developers started using agents, with smaller but consistent gains in testing discipline",
          "metric": "+76.8%",
          "quote": "We find one large agent-associated effect: a 76.8% increase in PR description length"
        }
      ],
      "techniques": [
        "longitudinal study with control developers",
        "aidev dataset"
      ],
      "caveats": "Drafted from the abstract only. Uses the same AIDev dataset as two existing MSR 2026 stories. Measures developer behaviour, not delivered software.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "yureki-lab-deploy-gate-autonomous-claude-code-agent-340-deploys",
      "title": "An autonomous Claude Code agent ships 340+ production deploys behind a three-stage gate it cannot edit",
      "url": "https://www.shippedbyagents.dev/stories/yureki-lab-deploy-gate-autonomous-claude-code-agent-340-deploys/",
      "source_url": "https://dev.to/yureki_lab/how-i-built-a-deploy-gate-so-my-autonomous-coding-agent-can-ship-to-prod-safely-1egb",
      "source_name": "dev.to (yureki_lab)",
      "source_type": "builder",
      "date": "2026-10-01",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Deploys since the gate",
          "value": "\"340+ production deploys with zero human-paged incidents\" over six months"
        },
        {
          "label": "Rollbacks",
          "value": "\"23 rollbacks into 23 closed test gaps\""
        },
        {
          "label": "Original incident",
          "value": "EU customers saw USD prices for 11 minutes"
        },
        {
          "label": "Canary",
          "value": "5% of traffic for 10 minutes, hard thresholds against a live baseline"
        },
        {
          "label": "Pre-flight failures",
          "value": "\"maybe one PR in twelve\", mostly under-declared blast radius"
        },
        {
          "label": "False breaches",
          "value": "4 in six months, all during one upstream payment provider incident"
        },
        {
          "label": "Cost per deploy",
          "value": "\"about 12 minutes of wall-clock time and a few cents of metrics queries\""
        },
        {
          "label": "Gate size",
          "value": "\"maybe 400 lines total\""
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "340+ production deploys over six months with zero human-paged incidents",
          "metric": "340+ deploys, 0 pages",
          "quote": "Six months later it has shipped 340+ production deploys with zero human-paged incidents."
        },
        {
          "claim": "The gate costs about 12 minutes of wall-clock time and a few cents per deploy",
          "metric": "~12 min per deploy",
          "quote": "The extra cost per deploy is about 12 minutes of wall-clock time and a few cents of metrics queries."
        }
      ],
      "techniques": [
        "verifier sub-agent approves merges",
        "canary with hard thresholds",
        "pre-flight checks out of the agent's reach",
        "automatic rollback"
      ],
      "caveats": "Self-reported by a single pseudonymous author on dev.to about one team and one system. The company, codebase size and traffic are not named, and none of the numbers are independently checked. The author says every incident the gate caught was a wrong answer (\"None of them produced a single 5xx\"); commenters point out a 5% canary over 10 minutes may miss low-volume regressions, and that the canary thresholds come from the deploy block the agent writes, so the agent still chooses how strict its own check is. No figures are given for the agent itself (token cost, time per task) or for how long the gate took to build.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "brian-erickson-gpxplore-ios-app-web-planner-agents-134-days-app-store",
      "title": "Product manager ships GPXplore, a motorcycle trip planner with a web app and native iPhone app, to the App Store without writing the code himself",
      "url": "https://www.shippedbyagents.dev/stories/brian-erickson-gpxplore-ios-app-web-planner-agents-134-days-app-store/",
      "source_url": "https://brian-erickson.com/blog/building-and-releasing-an-app-with-agents/",
      "source_name": "Brian Erickson (personal blog)",
      "source_type": "builder",
      "date": "2026-10-01",
      "tools": [
        "Claude Code",
        "OpenAI Codex",
        "Lovable"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Calendar time",
          "value": "134 days from first prototype commits to the 1 October App Store launch"
        },
        {
          "label": "Commits",
          "value": "1,821 across the three main repositories, on 100 distinct days"
        },
        {
          "label": "Merged pull requests",
          "value": "201"
        },
        {
          "label": "Code size",
          "value": "about 149,000 product lines and 79,000 test lines (including comments and blank lines)"
        },
        {
          "label": "Plans",
          "value": "entry-level paid plans for Claude and Codex"
        }
      ],
      "duration_days": 134,
      "cost_usd": null,
      "prs_merged": 201,
      "claimed_benefits": [
        {
          "claim": "An iOS app and web planner launched on the App Store 134 days after the first prototype commits, built on entry-level paid plans",
          "metric": "134 days"
        }
      ],
      "techniques": [
        "spec-driven development",
        "grilling session",
        "orchestrator alongside a spec",
        "switching between claude and codex when usage ran out"
      ],
      "caveats": "Self-reported by the builder. He says \"Claude\" and \"Codex\" without always naming the exact product, and doesn't remember the model versions. Line counts include comments and blanks. No cost total, user numbers or revenue. Only a few other people tested the app before launch, and he says much of the validation was his own use.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "julien-renaux-token-clicker-browser-game-claude-code-19-hours",
      "title": "Token Clicker: a 33,716-line browser idle game shipped 19 hours after its specs with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/julien-renaux-token-clicker-browser-game-claude-code-19-hours/",
      "source_url": "https://julienrenaux.fr/2026/09/28/token-clicker/",
      "source_name": "julienrenaux.fr",
      "source_type": "builder",
      "date": "2026-09-28",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Specs to launch",
          "value": "19 hours 39 minutes wall-clock (night included)"
        },
        {
          "label": "TypeScript in src/",
          "value": "33,716 lines"
        },
        {
          "label": "Tests",
          "value": "1,758 unit, 34 pacing, 28 end-to-end"
        },
        {
          "label": "Languages",
          "value": "7"
        },
        {
          "label": "Post-launch local cost",
          "value": "~$264 at API prices (24–28 Sept)"
        }
      ],
      "duration_days": 0.819,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Game shipped 19 hours and 39 minutes after the design specs were committed",
          "metric": "19 h 39 min",
          "quote": "The design specs were committed at 23:11 on 23 September, and the game shipped at 18:50 the next day: 19 hours and 39 minutes later."
        }
      ],
      "techniques": [
        "specs committed first",
        "agents in parallel",
        "test suite for the jokes",
        "denylist of real people and products"
      ],
      "caveats": "The 19 hours excludes writing the specs. Launch day ran in Claude Code on the web and isn't in the cost figure; the ~$264 covers only later local work. The game promotes the author's TrendingDevs brand.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "juno-posthog-self-driving-prs-26-of-last-30-roughly-90-percent-merge-rate",
      "title": "Juno merges 26 of its last 30 PostHog self-driving PRs, a roughly 90% merge rate, with the CEO still reviewing each one",
      "url": "https://www.shippedbyagents.dev/stories/juno-posthog-self-driving-prs-26-of-last-30-roughly-90-percent-merge-rate/",
      "source_url": "https://posthog.com/customers/juno",
      "source_name": "PostHog customer story",
      "source_type": "vendor",
      "date": "2026-09-23",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Self-driving PR merge rate",
          "value": "Roughly 90%"
        },
        {
          "label": "Recent self-driving PRs merged",
          "value": "26 of the last 30"
        },
        {
          "label": "Issues fixed by the same loop",
          "value": "Dozens, including a hung chat box, dose-logging errors on certain screen sizes, frozen tabs, and dead taps that lost drafts"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Juno CEO Marshall Gould says the merge rate for self-driving PRs is roughly 90%",
          "metric": "roughly 90%",
          "quote": "Our merge rate for self-driving PRs is roughly 90% now that we use PostHog for self-driving."
        },
        {
          "claim": "Gould says Juno has merged 26 of its last 30 self-driving PRs",
          "metric": "26 of the last 30",
          "quote": "Right now we've merged 26 of our last 30 self-driving PRs though, and we think it's going great."
        },
        {
          "claim": "Gould says the agents run autonomously and leave PRs ready to review and merge in the morning",
          "quote": "It's incredible that these agents can run autonomously to identify these things. You wake up in the morning and there are PRs waiting for you. All you have to do is review and merge."
        },
        {
          "claim": "PostHog says the same loop has fixed dozens of other issues, including a hung chat box, dose-logging errors on certain screen sizes, frozen tabs, and dead taps that lost drafts",
          "metric": "dozens of other issues",
          "quote": "The same loop has fixed dozens of other issues for Juno: a chat box that hung so people couldn't message the in-app AI, dose-logging errors that only appeared on certain screen sizes, users getting stuck on frozen tabs, and dead taps that were losing drafts in the symptom logging flow."
        }
      ],
      "techniques": [
        "agents open prs from analytics and error tracking",
        "human review on every pr",
        "scouts detect rage clicks without reading user data",
        "ai tests high-confidence fixes"
      ],
      "caveats": "Vendor-published PostHog customer story promoting PostHog self-driving, the product PostHog sells; the page has no author byline. Figures are self-reported by Juno CEO Marshall Gould. Roughly 90% and 26 of the last 30 describe a recent window, not a total volume, a time period, or a share of all Juno PRs. No calendar duration, cost or total merged-PR count is given. Gould still reads every PR, sometimes edits them, and often tests on a simulator; he has not turned on auto-merge. Some bugs cannot be reproduced easily because they depend on device state such as charging or low power mode. The page does not name the underlying coding model or a catalogue coding agent. \"Fixed dozens of other issues\" is PostHog's uncounted wording; the \"Log all\" medication-log example is described as Juno adding a feature after PostHog identified friction, without an explicit statement that self-driving authored that PR.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "github-copilot-runtime-typescript-to-rust",
      "title": "GitHub ports the Copilot agent runtime from TypeScript to 800K+ lines of Rust, mostly by one engineer",
      "url": "https://www.shippedbyagents.dev/stories/github-copilot-runtime-typescript-to-rust/",
      "source_url": "https://github.blog/ai-and-ml/generative-ai/migrating-the-github-copilot-runtime-to-rust-using-copilot/",
      "source_name": "The GitHub Blog",
      "source_type": "vendor",
      "date": "2026-09-16",
      "tools": [
        "GitHub Copilot"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Production Rust written",
          "value": "832,378 lines"
        },
        {
          "label": "TypeScript passed through the port (author estimate)",
          "value": "~430,000 lines"
        },
        {
          "label": "Porting window",
          "value": "~14.5 weeks"
        },
        {
          "label": "Port pull requests",
          "value": "128"
        },
        {
          "label": "Token spend",
          "value": "~$120,000 (~136.3B tokens)"
        },
        {
          "label": "Developer time (author estimate)",
          "value": "~3 weeks"
        },
        {
          "label": "Prompt-cache hit rate",
          "value": "96.22%"
        },
        {
          "label": "One-turn session, in-process vs before",
          "value": "5.25 s → 292 ms"
        }
      ],
      "duration_days": 101.5,
      "cost_usd": 120000,
      "prs_merged": 128,
      "claimed_benefits": [
        {
          "claim": "Runtime rewritten as 832,378 lines of production Rust for about $120,000 in tokens plus three weeks of a developer's time",
          "metric": "~$120,000 + 3 weeks",
          "quote": "the rough bill for the porting effort was about $120,000 in attributed token spend plus three weeks of a developer’s time"
        },
        {
          "claim": "A one-turn session went from 5.25 s to 292 ms in-process (18x)",
          "metric": "18x",
          "quote": "292 ms (18.0x)"
        },
        {
          "claim": "Ten-client memory delta cut by 91% before any optimisation work",
          "metric": "-91% memory",
          "quote": "cutting the measured ten-client memory delta by 91% before that optimization work begins"
        }
      ],
      "techniques": [
        "incremental component-by-component port",
        "shipping ports in pre-releases first",
        "compiler as fast feedback loop",
        "no big-bang cutover"
      ],
      "caveats": "Published by GitHub about its own product. The author notes team members also contributed (bindings, packaging, reviews), and that performance numbers include other changes from the same period.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "comarch-devin-clarion-erp-10-million-lines-5-years-to-18-months",
      "title": "Comarch compresses a five-year rebuild of its 10-million-line Clarion ERP to 18 months with Devin",
      "url": "https://www.shippedbyagents.dev/stories/comarch-devin-clarion-erp-10-million-lines-5-years-to-18-months/",
      "source_url": "https://devin.ai/customers/comarch",
      "source_name": "Devin customer story (Cognition)",
      "source_type": "vendor",
      "date": "2026-09-16",
      "tools": [
        "Devin"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "About 10 million lines of Clarion, 20 years old"
        },
        {
          "label": "Timeline",
          "value": "5-year plan compressed to 18 months (3.3x)"
        },
        {
          "label": "Epic cycle time",
          "value": "Median 80.5 days to 13.5 days (6x)"
        },
        {
          "label": "Merged PRs",
          "value": "2.8x more per month, year over year"
        },
        {
          "label": "Test coverage",
          "value": "60% to nearly 100%"
        },
        {
          "label": "Business analysis per feature",
          "value": "15+ man-days to 2-3 days"
        },
        {
          "label": "People onboarded",
          "value": "100"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A five-year modernisation compressed to 18 months",
          "metric": "5 years → 18 months",
          "quote": "a five-year modernization compressed to 18 months"
        },
        {
          "claim": "Business analysis per feature dropped from more than 15 man-days to 2 to 3 days, and median epic cycle time went from 80.5 days to 13.5",
          "metric": "6x faster epics",
          "quote": "Business analysis for new functionality dropped from more than 15 man-days to 2–3 days"
        },
        {
          "claim": "Monthly merged PRs up 2.8x year over year",
          "metric": "2.8x"
        },
        {
          "claim": "Automated test coverage up from 60% to nearly 100%",
          "metric": "60% → ~100%",
          "quote": "automated test coverage on the transformed system rose from 60% to nearly 100%"
        }
      ],
      "techniques": [
        "proof of concept against other vendors",
        "business logic extracted into user stories",
        "analysts review the stories",
        "agent generates and maintains tests"
      ],
      "caveats": "Published by Cognition as a customer story. The 18-month timeline is a plan in progress, not a finished migration. No cost or output code-size figures.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-09-16 from the page's JSON-LD datePublished."
    },
    {
      "slug": "julien-renaux-thumb-and-pylon-tower-defense-claude-code-2755-dollars",
      "title": "Solo developer builds a cross-platform tower defense game in about 30 days with Claude Code, for ~$2,755 in tokens",
      "url": "https://www.shippedbyagents.dev/stories/julien-renaux-thumb-and-pylon-tower-defense-claude-code-2755-dollars/",
      "source_url": "https://julienrenaux.fr/2026/09/10/thumb-and-pylon-vibe-coded/",
      "source_name": "julienrenaux.fr",
      "source_type": "builder",
      "date": "2026-09-10",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Duration",
          "value": "~30 days, ~12 working sessions"
        },
        {
          "label": "Tool calls",
          "value": "9,436"
        },
        {
          "label": "Output tokens",
          "value": "17.7M"
        },
        {
          "label": "Total tokens",
          "value": "9.24B (8.98B cache reads)"
        },
        {
          "label": "Cost at API prices",
          "value": "~$2,755 (~93% Claude Opus 5)"
        },
        {
          "label": "Levels",
          "value": "34"
        }
      ],
      "duration_days": 30,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A tower defense game for iOS and Android built from one design doc in about a month, for roughly $2,755 at API prices",
          "metric": "~30 days; ~$2,755",
          "quote": "the total came to roughly $2,755"
        }
      ],
      "techniques": [
        "single game design doc",
        "deterministic simulation shared with the server",
        "server-side re-simulation",
        "measuring with ccusage"
      ],
      "caveats": "Cost is the pay-as-you-go API-equivalent price from ccusage, not what he paid on a subscription. He says the token counts and dollar figure come from two tools that count cache reads differently. No code size given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "mistral-fortran-77-to-cpp-reservoir-simulator-40000-lines-parity-harness",
      "title": "Mistral migrates 40,000 lines of a Fortran 77 reservoir simulator to C++ for a European energy operator, using a numerical parity harness and human-run agent workflows",
      "url": "https://www.shippedbyagents.dev/stories/mistral-fortran-77-to-cpp-reservoir-simulator-40000-lines-parity-harness/",
      "source_url": "https://mistral.ai/news/legacy-code-modernization/",
      "source_name": "Mistral AI blog (Carlo Antonio Patti, Rasul Alakbarli)",
      "source_type": "vendor",
      "date": "2026-09-09",
      "tools": [
        "Mistral Vibe"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Lines migrated",
          "value": "40,000 of 300,000 lines of Fortran 77 (first sprint, core functionality)"
        },
        {
          "label": "Documentation agents",
          "value": "over a hundred"
        },
        {
          "label": "Module size",
          "value": "self-contained subtrees of under about 10,000 lines of Fortran"
        },
        {
          "label": "First autonomous attempt",
          "value": "one agent per subroutine over a week; functional but not modernised"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "40,000 lines of Fortran 77 migrated to C++ for a simulator with no test suite and no central documentation",
          "metric": "40,000 lines",
          "quote": "Mistral helped a European energy operator migrate 40,000 lines of Fortran 77 to C++"
        }
      ],
      "techniques": [
        "parity harness before migration",
        "agents document the codebase first",
        "caller-callee tree to find independent modules",
        "skill files to steer agents",
        "human oversight"
      ],
      "caveats": "Vendor post about its own models and services; the customer is anonymous. It gives no total duration, cost, or count of PRs, and Vibe CLI is named only for the documentation step, not explicitly for the migration agents. Only the first 40,000 lines were done.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "ultrathink-claude-code-vs-codex-30-day-fleet-completed-work",
      "title": "Ultrathink’s production agent fleet: after 30 days on Codex vs a frozen Claude Code baseline, Codex did not win on completion or QA",
      "url": "https://www.shippedbyagents.dev/stories/ultrathink-claude-code-vs-codex-30-day-fleet-completed-work/",
      "source_url": "https://ultrathink.art/blog/codex-vs-claude-code-30-days-completed-work",
      "source_name": "Ultrathink Field Notes",
      "source_type": "builder",
      "date": "2026-09-07",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Claude completed / terminal",
          "value": "1,286 / 1,317 (97.65%); started 1,321"
        },
        {
          "label": "Codex completed / terminal",
          "value": "574 / 631 (90.97%); started 656"
        },
        {
          "label": "Strict first-pass QA proxy",
          "value": "Claude 134/206 (65.05%); Codex 70/144 (48.61%)"
        },
        {
          "label": "Median wall time",
          "value": "Claude 422s; Codex 779s"
        },
        {
          "label": "p90 wall time",
          "value": "Claude 1,477s; Codex 3,126s"
        },
        {
          "label": "Windows",
          "value": "Claude ~2026-07-08 to 2026-08-07; Codex ~2026-08-07 to 2026-09-06"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Frozen Claude cohort completed a higher share of terminal tasks than the Codex period",
          "metric": "97.65% vs 90.97%",
          "quote": "Codex did not win."
        },
        {
          "claim": "Strict first-pass QA proxy favoured the Claude period by 16.44 points",
          "metric": "65.05% vs 48.61%"
        }
      ],
      "techniques": [
        "frozen measurement windows",
        "preserve raw denominators",
        "separate missing from zero",
        "attribute by executor at task start",
        "exclude migration burn-in interval"
      ],
      "caveats": "Self-reported by Ultrathink about its own fleet. Consecutive periods, not a controlled A/B; social vs operations workload mix shifted sharply. Missing cost and intervention metrics prevent a price-adjusted verdict. Not a claim that Claude Code is better for every team.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "coinbase-design-system-code-connect-full-coverage-four-hours-claude-code-22-percent-cheaper",
      "title": "Coinbase uses Claude Code to map its whole design system to Figma Code Connect in about four hours",
      "url": "https://www.shippedbyagents.dev/stories/coinbase-design-system-code-connect-full-coverage-four-hours-claude-code-22-percent-cheaper/",
      "source_url": "https://www.figma.com/blog/how-coinbase-used-code-connect-to-shrink-token-costs/",
      "source_name": "Figma blog",
      "source_type": "vendor",
      "date": "2026-09-02",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Sonnet 4.6"
      ],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Code Connect coverage",
          "value": "A few components to full design system in about 4 hours"
        },
        {
          "label": "Token use with Code Connect",
          "value": "-11.5% average"
        },
        {
          "label": "Time to implementation",
          "value": "-22.3% average"
        },
        {
          "label": "Cost",
          "value": "-22.5% average"
        },
        {
          "label": "Evaluation runs",
          "value": "3"
        }
      ],
      "duration_days": 0.167,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Full Code Connect coverage of the design system in about four hours",
          "metric": "~4 hours",
          "quote": "In about four hours, we went from having a few of our components migrated to the newer Code Connect format to full coverage of our design system"
        },
        {
          "claim": "Code Connect cut cost by 22.5%, time to implementation by 22.3% and token use by 11.5% on average",
          "metric": "-22.5% cost"
        },
        {
          "claim": "Agents stopped hallucinating illustration and icon names",
          "quote": "After completing our adoption of Code Connect, that problem disappeared overnight."
        }
      ],
      "techniques": [
        "parallel agentic workflow",
        "agent skills",
        "design-to-code context",
        "controlled evaluation runs"
      ],
      "caveats": "Published by Figma about its own Code Connect product. The evaluation was three runs on one contrived design. Exact component count is given only as hundreds.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "carlos-infantes-skillpilot-paid-product-claude-code-three-weeks-1805-tests-missed-checkout-bug",
      "title": "Solo builder ships SkillPilot, a paid product, with Claude Code in about three weeks, but 1,805 passing tests missed a dead checkout",
      "url": "https://www.shippedbyagents.dev/stories/carlos-infantes-skillpilot-paid-product-claude-code-three-weeks-1805-tests-missed-checkout-bug/",
      "source_url": "https://dev.to/carlosinfantes/i-shipped-a-paid-product-with-claude-code-1805-green-tests-didnt-catch-the-bug-that-mattered-bml",
      "source_name": "Carlos Infantes (DEV Community)",
      "source_type": "builder",
      "date": "2026-09-01",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Build time",
          "value": "About three weeks"
        },
        {
          "label": "Tests",
          "value": "1,805 unit tests in 103 test files against 306 source files"
        },
        {
          "label": "Production bugs missed by tests",
          "value": "3 at first purchase (CSP-blocked checkout, over-broad access, wrong price), then a dead analytics tracker"
        },
        {
          "label": "Scope",
          "value": "Two Next.js apps, 24 sellable items"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "The first 80% of the product (schema, auth, checkout, webhook, entitlements) took days, not weeks",
          "quote": "Schema, auth, content pipeline, checkout, webhook, entitlements — days, not weeks."
        }
      ],
      "techniques": [
        "state design decisions before prompting",
        "design skill with concrete site references",
        "workflow skills",
        "atomic database function for coupon limits",
        "end-to-end test with a real card"
      ],
      "caveats": "Self-reported by the builder. No revenue, user or exact token cost figures; the cost claim is relative to his previous tooling. Mostly a lessons post rather than an outcome report.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "group-software-vb6-erp-claude-code-12-features-70-percent-equivalence",
      "title": "Industrial case study: Claude Code migrates 12 VB6 ERP features to C# .NET 10 at 70% equivalence for $50.29 in about an hour",
      "url": "https://www.shippedbyagents.dev/stories/group-software-vb6-erp-claude-code-12-features-70-percent-equivalence/",
      "source_url": "https://arxiv.org/abs/2608.28972",
      "source_name": "arXiv 2608.28972 (Alves, Politowski, Montandon)",
      "source_type": "independent",
      "date": "2026-08-29",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 4.6"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Features migrated",
          "value": "12 (6 low, 4 medium, 2 high complexity)"
        },
        {
          "label": "Overall equivalence",
          "value": "70% of 331 instructions (persistence 71%, functional 69%)"
        },
        {
          "label": "Equivalence by complexity",
          "value": "Low 92%, medium 81%, high 47%"
        },
        {
          "label": "Total agent time",
          "value": "59 min 45 s (avg 4 min 54 s per feature)"
        },
        {
          "label": "Tokens and cost",
          "value": "45.8M tokens, $50.29 total (avg $4.19 / 3.81M tokens per feature)"
        },
        {
          "label": "Legacy modules",
          "value": "Measurements ~299k LOC (219 clients); Purchases ~224k LOC (61 clients)"
        },
        {
          "label": "Inter-rater agreement on rules",
          "value": "Cohen’s κ 0.84; 93% agreement on 160 rules"
        }
      ],
      "duration_days": 0.041,
      "cost_usd": 50.29,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "70% average equivalence across 331 evaluated instructions",
          "metric": "70%"
        },
        {
          "claim": "Low-complexity features reached 92% equivalence at $1.66 average",
          "metric": "92% / $1.66"
        },
        {
          "claim": "All 12 features migrated in under an hour of agent time for $50.29",
          "metric": "59m 45s / $50.29"
        }
      ],
      "techniques": [
        "single-pass generation",
        "claude.md per feature environment",
        "persistence and functional parity scoring",
        "complexity banding by loc methods dependencies",
        "read-only legacy directory"
      ],
      "caveats": "Single-pass generation with no human refinement between iterations, so results understate iterative workflows. One VB6 ERP and one agent/model configuration. Only two high-complexity features.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-08-29 from the arXiv submission date on abs/2608.28972."
    },
    {
      "slug": "uber-software-factory-70-percent-prs-agents-cost-per-session-down-52-percent",
      "title": "Uber says agents are behind more than 70% of its pull requests, and explains how it cut cost per session 52% while usage grew 7x",
      "url": "https://www.shippedbyagents.dev/stories/uber-software-factory-70-percent-prs-agents-cost-per-session-down-52-percent/",
      "source_url": "https://www.uber.com/us/en/blog/efficient-software-factory/",
      "source_name": "Uber Engineering blog (Uday Kiran Medisetty)",
      "source_type": "builder",
      "date": "2026-08-27",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Agent-attributed PRs",
          "value": "More than 70% of pull requests"
        },
        {
          "label": "Agent skills",
          "value": "Over 3,600 built; more than 30K executions per day"
        },
        {
          "label": "Growth Feb to Aug 2026",
          "value": "Weekly active users 7x, weekly agent requests 9.4x"
        },
        {
          "label": "Cost with model held fixed",
          "value": "Cost per 1,000 requests down almost 34% from peak; cost per session down 52% from June peak"
        },
        {
          "label": "Code-mode token savings",
          "value": "55-71% fewer tokens on small SQL queries"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 70% of pull requests are attributed to local or cloud agents",
          "metric": "70%+",
          "quote": "More than 70% of pull requests are attributed to local or cloud agents."
        },
        {
          "claim": "With the model held fixed, cost per 1,000 requests down almost 34% from peak and cost per session down 52% from the June peak",
          "metric": "-34% / -52%",
          "quote": "cost per session is down 52% from its June peak"
        },
        {
          "claim": "Code-mode cut tokens by 55% to 71% on small SQL queries",
          "metric": "55–71% fewer tokens"
        }
      ],
      "techniques": [
        "managed agents",
        "benchmark-driven model selection",
        "cheaper model for subagents",
        "code-mode batching of tool calls",
        "context graph"
      ],
      "caveats": "Company writing about its own platform. 'Attributed to agents' isn't defined in the post. Mostly about cost and tooling rather than one delivered project. A related Pragmatic Engineer interview says the background agent Minion writes about 15% of production PRs, with a person reviewing every PR.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "doordash-multi-agent-feature-flag-cleanup-45-of-50-prs-4-79-dollars-14-minutes",
      "title": "DoorDash's multi-agent system turns stale feature flags into merge-ready PRs for $4.79 and 14 minutes each",
      "url": "https://www.shippedbyagents.dev/stories/doordash-multi-agent-feature-flag-cleanup-45-of-50-prs-4-79-dollars-14-minutes/",
      "source_url": "https://careersatdoordash.com/blog/automating-feature-flag-cleanup-at-scale-with-a-multi-agent-llm-system/",
      "source_name": "DoorDash Engineering blog",
      "source_type": "builder",
      "date": "2026-08-24",
      "tools": [
        "Claude"
      ],
      "models": [
        "Claude Sonnet",
        "Claude Opus"
      ],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Usable PRs",
          "value": "45 of 50 stale flags"
        },
        {
          "label": "Merged first shot",
          "value": "31 of 50"
        },
        {
          "label": "Average cost per flag",
          "value": "$4.79"
        },
        {
          "label": "Average time per flag",
          "value": "13.8 minutes (vs 1-2 hours manually)"
        },
        {
          "label": "Bugs or regressions introduced",
          "value": "None observed across 50 PRs"
        },
        {
          "label": "Scale",
          "value": "60,000+ flags across ~623 repositories"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Usable PRs for 45 of the 50 most recent stale flags",
          "metric": "45/50",
          "quote": "the system produced usable PRs for 45 of them at an average cost of $4.79 and 13.8 minutes per flag"
        },
        {
          "claim": "13.8 minutes per cleanup instead of one to two hours manually",
          "metric": "13.8 minutes"
        },
        {
          "claim": "No bugs or regressions introduced across all 50 generated PRs",
          "quote": "The system introduced no bugs or regressions along the way."
        },
        {
          "claim": "Using Sonnet for orchestration and Opus only for editing cut overall API cost"
        }
      ],
      "techniques": [
        "orchestrator plus parallel removal agents",
        "git worktree isolation",
        "live rollout state over mcp",
        "human checkpoint before edits",
        "deterministic validation gates"
      ],
      "caveats": "Self-reported by the team that built it, on 50 flags in Kotlin repositories. A human still confirms each target value and reviews the PR. Complex flags reached 85% (22/26) and outliers took about 40 minutes and $19.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "asana-codex-enzyme-to-rtl-two-weeks",
      "title": "Asana finishes a five-year Enzyme → React Testing Library migration in two weeks with Codex",
      "url": "https://www.shippedbyagents.dev/stories/asana-codex-enzyme-to-rtl-two-weeks/",
      "source_url": "https://openai.com/index/asana/",
      "source_name": "OpenAI customer story (Asana also published its own account, in partnership with OpenAI)",
      "source_type": "vendor",
      "date": "2026-08-19",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Test migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Calendar time",
          "value": "~2 weeks"
        },
        {
          "label": "Engineering effort",
          "value": "~1.5 weeks"
        },
        {
          "label": "Parallel agents",
          "value": "Up to 4"
        },
        {
          "label": "Model + infrastructure cost",
          "value": "~$12K (model ~$11K, infra ~$1K)"
        },
        {
          "label": "Projected time at prior pace",
          "value": "~5 more years"
        },
        {
          "label": "Estimated manual cost",
          "value": "~$6M fully loaded (Asana estimate)"
        }
      ],
      "duration_days": 14,
      "cost_usd": 12000,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Migration finished in about two weeks, where the prior pace projected roughly five more years",
          "metric": "~2 weeks vs ~5 years",
          "quote": "With Codex, Asana completed the project in about two weeks."
        },
        {
          "claim": "About $12K in model and infrastructure costs, against Asana's estimate of roughly $6M in manual engineering effort",
          "metric": "~$12K vs ~$6M",
          "quote": "would have represented roughly $6M in fully loaded engineering effort"
        },
        {
          "claim": "Improved test coverage, fixed poor-quality tests and cleaned up legacy testing infrastructure",
          "quote": "improving test coverage, fixing poor-quality tests, and cleaning up legacy testing infrastructure"
        }
      ],
      "techniques": [
        "five-sentence prompt",
        "up to four parallel agents",
        "separate copy of the codebase per agent",
        "engineer reviewed every change",
        "simpler instructions worked better"
      ],
      "caveats": "Published by OpenAI; Asana's own post notes it is part of a partnership with OpenAI. The $6M figure is Asana's back-of-the-napkin estimate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "zalando-agentic-engineering-snapshot-risk-based-pr-auto-approval-33-percent",
      "title": "Zalando's 2.5 years of agentic engineering across 250+ teams: a risk-based bot auto-approves 33% of PRs and cuts their lead time 20-40%",
      "url": "https://www.shippedbyagents.dev/stories/zalando-agentic-engineering-snapshot-risk-based-pr-auto-approval-33-percent/",
      "source_url": "https://engineering.zalando.com/posts/2026/08/agentic-engineering-at-zalando-a-snapshot.html",
      "source_name": "Zalando Engineering Blog (Bartosz Ocytko)",
      "source_type": "builder",
      "date": "2026-08-14",
      "tools": [
        "GitHub Copilot",
        "Claude Code",
        "OpenCode"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Teams",
          "value": "More than 250 engineering teams"
        },
        {
          "label": "Auto-approved PRs",
          "value": "33% rated low risk and auto-approved"
        },
        {
          "label": "PR lead time",
          "value": "Reduced 20-40% for auto-approved PRs"
        },
        {
          "label": "LLM proxy",
          "value": "About 2k MAU on six small pods"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A risk-based bot auto-approves 33% of PRs",
          "metric": "33%",
          "quote": "33% of our PRs are low-risk and are auto-approved by the bot"
        },
        {
          "claim": "Auto-approval reduced PR lead time by 20-40%",
          "metric": "20-40%",
          "quote": "reduced PR lead time by 20-40% (when compared with all PRs)"
        }
      ],
      "techniques": [
        "llm proxy for all model access",
        "risk-based pr auto-approval",
        "central agent skill collection",
        "migration skills",
        "fleet-wide agent transformations"
      ],
      "caveats": "Written by Zalando's own engineering leadership. The 20-40% lead-time cut is for auto-approved PRs compared with all PRs. No productivity or cost totals; complexity findings come from four example codebases. Also mentions pi, a coding agent not in the site's tool list.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "linkedin-multi-agent-ai-code-review-79000-reviews-a-week-63-9-percent-accepted",
      "title": "LinkedIn's multi-agent AI code reviewer runs 79,000 reviews a week with 63.9% of suggestions accepted",
      "url": "https://www.shippedbyagents.dev/stories/linkedin-multi-agent-ai-code-review-79000-reviews-a-week-63-9-percent-accepted/",
      "source_url": "https://www.linkedin.com/blog/engineering/ai/high-signal-ai-code-review-that-adapts-to-your-codebase-at-scale",
      "source_name": "LinkedIn Engineering blog (Min Chen et al.)",
      "source_type": "builder",
      "date": "2026-08-13",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Weekly reviews",
          "value": "79,000+ across 40,000+ PRs in 7,500+ repositories"
        },
        {
          "label": "Repositories adopted",
          "value": "~10,000"
        },
        {
          "label": "Combined acceptance rate",
          "value": "63.9% (sampled)"
        },
        {
          "label": "Task completion rate",
          "value": "99.1%"
        },
        {
          "label": "Review latency",
          "value": "Starts within ~90 s, most finish under 10 minutes"
        },
        {
          "label": "Acceptance by category",
          "value": "Logic errors 80%, bug fixes 58.1%, refactoring 43.5%, security 40.6%"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "63.9% combined acceptance rate for automated review suggestions on sampled PRs",
          "metric": "63.9%"
        },
        {
          "claim": "Most reviews are posted before the first human reviewer opens the PR",
          "quote": "Most reviews are posted before the first human reviewer opens the PR."
        }
      ],
      "techniques": [
        "multi-agent review",
        "cross-model deduplication as signal",
        "three-tier custom rules",
        "pairwise llm-judge eval",
        "post-merge acceptance eval"
      ],
      "caveats": "Builder-written. The models and harnesses behind the subagents are not named. This is review of code, not agent-written code; acceptance was judged by an LLM evaluator on a 7-day sample.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "abenhaim-717k-line-typescript-refactor-189-files-no-human-review-2430-dollars",
      "title": "An AI coding agent rips out a core invariant across 189 files of a 717,000-line TypeScript app in three days, with no human code review",
      "url": "https://www.shippedbyagents.dev/stories/abenhaim-717k-line-typescript-refactor-189-files-no-human-review-2430-dollars/",
      "source_url": "https://arxiv.org/abs/2608.12440",
      "source_name": "arXiv 2608.12440 (Joel Abenhaim)",
      "source_type": "builder",
      "date": "2026-08-12",
      "tools": [
        "AICode"
      ],
      "models": [
        "OpenAI ChatGPT 5.6 Sol"
      ],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "717,725 lines of TypeScript, 3,648 files"
        },
        {
          "label": "Change",
          "value": "189 files touched (31 new); 288 files, +34,770 / -16,422 across both commits"
        },
        {
          "label": "Time",
          "value": "Three days"
        },
        {
          "label": "Cost",
          "value": "USD 2,430 in model inference"
        },
        {
          "label": "Audit passes",
          "value": "31 (14 spec refinement + 17 verification)"
        },
        {
          "label": "Defects fixed before first human run",
          "value": "201"
        }
      ],
      "duration_days": 3,
      "cost_usd": 2430,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "189 files changed in a 717,725-line codebase in three days for USD 2,430 in inference",
          "metric": "3 days; USD 2,430",
          "quote": "Elapsed: three days; cost: USD 2,430."
        },
        {
          "claim": "201 defects corrected across 31 audit passes before any human ran the program",
          "metric": "201 defects",
          "quote": "Across 31 audit passes, 201 defects were corrected before any human executed the program."
        }
      ],
      "techniques": [
        "specification-first protocol",
        "agent-written spec",
        "repeated audit passes",
        "second agent checks the first",
        "agent proposes patches"
      ],
      "caveats": "Self-reported: the author built the codebase, ran the operation and wrote the paper, and his company (AI Sovereign Labs) designs and distributes the agent used, so this is partly vendor. Single case; the codebase is closed source, though the build and raw logs (in French) are published.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "vercel-ai-sdk-factory-agents-author-25-35-percent-of-merged-prs-close-75-percent-of-issues",
      "title": "Vercel's AI SDK software factory authors 25–35% of merged PRs and closes most issues, four weeks into production",
      "url": "https://www.shippedbyagents.dev/stories/vercel-ai-sdk-factory-agents-author-25-35-percent-of-merged-prs-close-75-percent-of-issues/",
      "source_url": "https://vercel.com/blog/building-a-software-factory-for-ai-sdk",
      "source_name": "Vercel blog (Lars Grammel, Eric Dodds)",
      "source_type": "vendor",
      "date": "2026-08-12",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of weekly merged PRs authored by the factory",
          "value": "25–35%"
        },
        {
          "label": "Share of weekly merges to the v6 release line",
          "value": "Above 50% (v5 similar)"
        },
        {
          "label": "Closed issues closed by the factory",
          "value": "Over 75% in July"
        },
        {
          "label": "Open issues",
          "value": "1,022 (late June peak) → 844 (early August)"
        },
        {
          "label": "Open bugs",
          "value": "Down roughly 25%"
        },
        {
          "label": "Time in production",
          "value": "Just over four weeks"
        },
        {
          "label": "Human review",
          "value": "A human merges every change"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Factory agents author 25–35% of the PRs merged each week",
          "metric": "25–35% of weekly merged PRs",
          "quote": "Between 25-35% of the PRs we merge on a weekly basis are now authored by `ai-sdk-factory` agents."
        },
        {
          "claim": "Factory PRs are above 50% of weekly merges to the v6 release line",
          "metric": "Above 50% of v6 merges",
          "quote": "Factory PRs are above 50% of weekly merges to the v6 release line, and v5 looks similar."
        },
        {
          "claim": "Over 75% of closed issues in July were closed by the factory",
          "metric": "Over 75% of closed issues",
          "quote": "In July, over 75% of closed issues were closed by the factory."
        },
        {
          "claim": "Open issues fell from 1,022 to 844 and open bugs are down about 25%",
          "metric": "1,022 → 844 open issues",
          "quote": "Open issues fell from a peak of 1,022 in late June to 844 by early August, and open bugs are down roughly 25%."
        }
      ],
      "techniques": [
        "one agent per task",
        "isolated sandbox per agent",
        "build value locally, then move to the cloud",
        "evidence chain for human reviewer",
        "failed runs become new evals"
      ],
      "caveats": "Self-reported by Vercel, which sells the Sandbox, Functions, Queues and Blob products the factory runs on and promotes a one-click deployable version, so this is vendor-published. The post does not name which coding agent or model powers the factory's agents (it mentions Opus 4.6 only as the trigger for the PR surge), so the tool is Unstated. Results cover just over four weeks. The intro says the factory closes 70-80% of issues while the results section says over 75% of closed issues in July; these are shares of closed issues, not of all open issues. No absolute PR counts are given. The open-issue drop is not attributed solely to the factory.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "doordash-flux-cloud-agents-130000-tasks-a-month-25000-code-reviews-a-week",
      "title": "DoorDash's in-house Flux cloud agent platform automates 130,000 engineering tasks in a month",
      "url": "https://www.shippedbyagents.dev/stories/doordash-flux-cloud-agents-130000-tasks-a-month-25000-code-reviews-a-week/",
      "source_url": "https://careersatdoordash.com/blog/delegating-engineering-work-to-cloud-based-agents/",
      "source_name": "DoorDash Engineering blog",
      "source_type": "builder",
      "date": "2026-08-11",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Engineering tasks automated",
          "value": "130,000 in a single month"
        },
        {
          "label": "Automated code reviews",
          "value": "25,000+ per week"
        },
        {
          "label": "Playbooks",
          "value": "300+ unique, 10,000+ invocations per week"
        },
        {
          "label": "Sandbox setup",
          "value": "p95 SLO under 5 seconds"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Automated 130,000 engineering tasks in a single month",
          "metric": "130,000 tasks/month",
          "quote": "In a single month in 2026, we used Flux to automate 130,000 engineering tasks."
        },
        {
          "claim": "More than 25,000 automated code reviews each week",
          "metric": "25,000+/week"
        }
      ],
      "techniques": [
        "cloud sandboxes",
        "scoped mcp gateway",
        "reusable playbooks",
        "start narrow with code review",
        "public slack threads"
      ],
      "caveats": "Platform post from the builders. Which coding agent harnesses Flux runs is not named. Counts are usage volumes; no merge rates, quality or time-saved figures are given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "assembled-auto-approval-low-risk-prs-2-4x-prs-merged-43-percent-auto-approved",
      "title": "Assembled auto-approves low-risk PRs with Codex and Claude Code reviewers, lifting PRs merged to 2.4× its pre-agent baseline",
      "url": "https://www.shippedbyagents.dev/stories/assembled-auto-approval-low-risk-prs-2-4x-prs-merged-43-percent-auto-approved/",
      "source_url": "https://www.assembled.com/blog/code-review-bottlenecks-how-we-hill-climbed-our-way-to-higher-pr-throughput",
      "source_name": "Assembled blog (John Wang, co-founder and CTO)",
      "source_type": "builder",
      "date": "2026-08-07",
      "tools": [
        "OpenAI Codex",
        "Claude Code",
        "Devin"
      ],
      "models": [
        "GPT-5.6-sol",
        "Claude Opus 5"
      ],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs merged",
          "value": "2.4× pre-agent baseline"
        },
        {
          "label": "Lines of code",
          "value": "2.8× pre-agent baseline"
        },
        {
          "label": "Auto-approvals",
          "value": "43% of merges, three weeks after full-team launch"
        },
        {
          "label": "Large PRs (600+ lines) throughput",
          "value": "3.5×, p90 wait time steady"
        },
        {
          "label": "Refactoring / test work",
          "value": "7.5× / 3.4×"
        },
        {
          "label": "Bugs reported per week",
          "value": "Down 27%"
        },
        {
          "label": "Reverts per PR",
          "value": "Steady at 0.8%"
        },
        {
          "label": "Median auto-review time",
          "value": "8 minutes"
        },
        {
          "label": "Estimated LLM cost",
          "value": "$3.57 per review attempt (14-day sample)"
        },
        {
          "label": "Time to first review before auto-review",
          "value": "3.5 hours (Dec 2025) → over 16 hours (May 2026)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "PRs merged rose to 2.4× the pre-agent baseline",
          "metric": "2.4× PRs merged",
          "quote": "PRs merged rose to 2.4× of our pre-agent baseline."
        },
        {
          "claim": "Auto-approvals grew to 43% of merges within three weeks of full-team launch",
          "metric": "43% of merges",
          "quote": "Auto-approvals grew to 43% of merges over the three weeks after the full-team launch."
        },
        {
          "claim": "Large PRs (600+ lines) rose 3.5× in throughput with p90 wait time steady",
          "metric": "3.5× large-PR throughput",
          "quote": "Large PRs (600+ lines) increased 3.5× in throughput, with p90 wait time holding steady."
        },
        {
          "claim": "Refactoring up 7.5× and test work up 3.4×",
          "metric": "7.5× refactoring, 3.4× tests",
          "quote": "Refactoring went up 7.5× and test work went up 3.4×."
        },
        {
          "claim": "Bugs reported per week down 27%, reverts steady at 0.8%",
          "metric": "-27% bugs reported",
          "quote": "Bugs reported per week went down 27% and reverts per PR stayed steady at 0.8%."
        }
      ],
      "techniques": [
        "two independent agent reviews",
        "orchestrator applies approval policy",
        "auto-approve low-risk prs",
        "testing evidence in pr description",
        "trusted alpha testers first"
      ],
      "caveats": "Self-reported by Assembled's CTO. The author calls these observational rollout results, not a randomized trial, and says the auto-reviewer doesn't get credit for every point. The window is short (three weeks after full-team launch), and the bug figures are explicitly too early to be conclusive. The pre-agent baseline period is not precisely dated, and no absolute PR counts are given. Category multipliers differ between the intro (bugfixes 3.7×, refactors 7.5×, features 2.5×) and the results list (refactoring 7.5×, test work 3.4×). The $3.57 per attempt is an estimate, not total spend. Codex and Claude Code are the reviewers; Devin is mentioned only for non-engineer PRs, and the agents that wrote most of the PRs are not broken down.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "ambrook-mui-to-base-ui-design-system-migration-agent-graph-794-files-two-weeks",
      "title": "Ambrook migrates its design system from Material UI to Base UI in just over two weeks with a graph of planner and worker agents",
      "url": "https://www.shippedbyagents.dev/stories/ambrook-mui-to-base-ui-design-system-migration-agent-graph-794-files-two-weeks/",
      "source_url": "https://ambrook.com/engineering/ai/migrating-fast-with-agent-graphs-design-system-material-ui-base-ui",
      "source_name": "Ambrook Engineering (Dan Schlosser, Adam Markon)",
      "source_type": "builder",
      "date": "2026-08-06",
      "tools": [
        "Claude Code",
        "Niteshift"
      ],
      "models": [
        "Claude Fable",
        "Claude Opus",
        "Claude Sonnet"
      ],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Change",
          "value": "54,000 lines changed across 794 files"
        },
        {
          "label": "Time",
          "value": "Just over two weeks from first commit to production"
        },
        {
          "label": "PRs and sessions",
          "value": "60 PRs merged across 121 agent sessions"
        },
        {
          "label": "Tokens",
          "value": "4.5B total (4.32B of them cache reads)"
        },
        {
          "label": "Tests added first",
          "value": "About 11,000 lines, roughly 1,270 cases, in nine PRs"
        },
        {
          "label": "Bug bash",
          "value": "Two days end-to-end"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 60,
      "claimed_benefits": [
        {
          "claim": "Migration of 54k lines across 794 files completed in just over two weeks",
          "metric": "just over two weeks",
          "quote": "we completed it in just over two weeks without any major regressions or customer disruption"
        },
        {
          "claim": "Bug bash and fixes cut to two days end-to-end",
          "metric": "two days"
        },
        {
          "claim": "Rebuilt components merged three days after the stack was opened",
          "metric": "three days"
        },
        {
          "claim": "Replaced MUI without major regressions",
          "quote": "we had replaced MUI without any major regressions"
        }
      ],
      "techniques": [
        "long-lived planner agent",
        "per-batch briefs as linear tickets",
        "tests before migration",
        "codemod for interface change",
        "adversarial review agents from different providers"
      ],
      "caveats": "Written by the team that did the work. No cost in dollars is given. 'Without any major regressions' is their own assessment. Niteshift and Supercut are third-party tools they used; the post does not say whether Ambrook has a commercial relationship with them.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "giles-edwards-alexander-150k-line-agent-built-app-refactor-cuts-tokens-83-percent",
      "title": "An agent-written 150,000-line Rust app shows refactoring one 17,155-line file cut input tokens per change by 83%",
      "url": "https://www.shippedbyagents.dev/stories/giles-edwards-alexander-150k-line-agent-built-app-refactor-cuts-tokens-83-percent/",
      "source_url": "https://martinfowler.com/articles/exploring-gen-ai/refactoring-economic-benefit.html",
      "source_name": "martinfowler.com (Exploring Gen AI series)",
      "source_type": "builder",
      "date": "2026-07-30",
      "tools": [
        "Claude Code",
        "Cursor"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "App size",
          "value": "~150,000 lines (~120k Rust)"
        },
        {
          "label": "Largest file before",
          "value": "17,155 lines"
        },
        {
          "label": "Largest file after",
          "value": "3,695 lines"
        },
        {
          "label": "Input tokens per change",
          "value": "159,564 → 27,360 (−83%)"
        },
        {
          "label": "Saving per change",
          "value": "~39.7 cents at $3/MTok (author's estimate)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Input tokens for the same task fell from 159,564 to 27,360 after refactoring, an 83% saving",
          "metric": "-83% tokens",
          "quote": "Between the base line and the final refactoring, input tokens for the same task reduced from 159,564 to 27,360."
        },
        {
          "claim": "Largest file cut from 17,155 lines to 3,695",
          "metric": "17,155 → 3,695 lines"
        }
      ],
      "techniques": [
        "refactoring for agent context",
        "fowler refactoring catalogue",
        "measuring tokens per change",
        "no human code review"
      ],
      "caveats": "One app and one repeated change, measured by the builder. The author notes the dollar saving per change is small and the wider effect is unproven. Build time and total cost aren't given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "brex-micronaut-4-migration-180-kotlin-services-one-engineer-one-month-cursor-agents",
      "title": "Brex migrates 180 Kotlin microservices to Micronaut 4 in a month with one engineer and two self-hosted Cursor agents, after a manual attempt managed 50 in a quarter",
      "url": "https://www.shippedbyagents.dev/stories/brex-micronaut-4-migration-180-kotlin-services-one-engineer-one-month-cursor-agents/",
      "source_url": "https://www.brex.com/journal/agent-automations-for-big-migrations",
      "source_name": "Brex Journal (engineering)",
      "source_type": "builder",
      "date": "2026-07-22",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Services migrated",
          "value": "180 microservices in one month by one engineer"
        },
        {
          "label": "Previous manual attempt",
          "value": "About 50 services in a quarter with three full-time engineers"
        },
        {
          "label": "Speed-up",
          "value": "32x per capita (Brex figure)"
        },
        {
          "label": "Cost per service migration",
          "value": "Down about 90% including infrastructure and token costs"
        },
        {
          "label": "Skills",
          "value": "Eleven, each scoped to one class of failure"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "One engineer migrated 180 microservices in a month, where three engineers had managed about 50 in a quarter by hand: a 32x per capita speed-up",
          "metric": "32x",
          "quote": "This represents a 32x per capita acceleration in delivery speed."
        },
        {
          "claim": "Cost per microservice migration down about 90%, counting infrastructure and tokens",
          "metric": "-90%",
          "quote": "we decreased our cost per microservice migration by approximately 90%"
        }
      ],
      "techniques": [
        "scripts plus modular skills",
        "agent triggered by linear tickets",
        "skills loaded only for the error at hand",
        "cloud agents instead of laptops"
      ],
      "caveats": "Brex's own blog, written up as a Cursor success. A few automated PRs passed all tests and still caused subtle runtime bugs in production. No absolute dollar costs, model names or PR counts. Published July 2026, so older than this week's scan.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "wix-code-migration-platform-agent-prs-650-developer-days-one-quarter-88-percent-merged-unchanged",
      "title": "Wix's Code Migration Platform has a coding agent open migration PRs across hundreds of repos, saving 650 developer-days in a quarter",
      "url": "https://www.shippedbyagents.dev/stories/wix-code-migration-platform-agent-prs-650-developer-days-one-quarter-88-percent-merged-unchanged/",
      "source_url": "https://www.wix.engineering/posts/how-wix-saved-650-developer-days-in-one-quarter-by-automating-code-migrations",
      "source_name": "Wix Engineering (Dana Avital)",
      "source_type": "builder",
      "date": "2026-07-22",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Developer-days saved",
          "value": "650 in a single quarter (their estimate)"
        },
        {
          "label": "PRs merged without manual changes",
          "value": "88%"
        },
        {
          "label": "Migrations fully merged within two weeks of PR creation",
          "value": "60%"
        },
        {
          "label": "Scale",
          "value": "About 50 migrations a year, 6,000 repositories touched a year"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "650 developer-days saved in a single quarter",
          "metric": "650 developer-days",
          "quote": "650 developer-days saved in a single quarter"
        },
        {
          "claim": "88% of PRs merged without manual changes",
          "metric": "88%"
        },
        {
          "claim": "60% of migrations fully merged within two weeks of PR creation",
          "metric": "60%"
        }
      ],
      "techniques": [
        "prompt bank under version control",
        "feed owner edits back into prompt",
        "automerge when safe",
        "single owner per change",
        "context in pr description"
      ],
      "caveats": "Self-reported. The days-saved figure is calculated from services affected times estimated manual time per service, compared with actual time. The post does not name the coding agent or models, and says scripts handle a significant share of migrations without an LLM.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "orb5x-gyrokinetic-plasma-code-fortran-to-kokkos-cursor-claude-code-codex-opencode",
      "title": "PSI, EPFL and Max Planck researchers translate the ORB5 plasma simulation code from Fortran to C++/Kokkos with Cursor, Claude Code, Codex and OpenCode, matching the original to near machine accuracy",
      "url": "https://www.shippedbyagents.dev/stories/orb5x-gyrokinetic-plasma-code-fortran-to-kokkos-cursor-claude-code-codex-opencode/",
      "source_url": "https://arxiv.org/abs/2607.17363",
      "source_name": "arXiv (Sadr, Lanti, Mishchenko, Wang, Villard)",
      "source_type": "builder",
      "date": "2026-07-19",
      "tools": [
        "Cursor",
        "Claude Code",
        "OpenCode",
        "OpenAI Codex"
      ],
      "models": [
        "qwen2.5-coder:32b",
        "gpt-oss:120b"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Benchmark agreement with Fortran ORB5",
          "value": "Almost machine accuracy for ITG, ITPA and chirping field components"
        },
        {
          "label": "Platforms",
          "value": "Linux and Mac laptops; LUMI-G, Daint-ALPS, CINECA, Discoverer"
        },
        {
          "label": "Scaling",
          "value": "Near-ideal on AMD MI250X (LUMI-G) and NVIDIA GH200 (Daint)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "ITG, ITPA and chirping tests agree with the Fortran original to almost machine accuracy"
        },
        {
          "claim": "Compiles and runs on personal laptops as well as several HPC clusters, lowering the cost of maintaining the Fortran code"
        }
      ],
      "techniques": [
        "fortran and c++ side by side in one tree",
        "hdf5 parity gate per patch",
        "smallest coherent patch per campaign",
        "cpu then multi-rank then gpu gates",
        "agent reports uncertainty and risks"
      ],
      "caveats": "Written by the researchers who did the translation. The paper gives no code size, duration, cost or share of code written by agents. The tuned Fortran ORB5 is still faster than ORB5X where it runs; the authors present ORB5X as an initial implementation, and some routines still stage through host memory for FFTW, HDF5, LAPACK and MPI. The strong-flow path should not be treated as production-ready until a rotating-plasma case is validated.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "anthropic-mike-krieger-python-to-typescript-165k-weekend-claude-code",
      "title": "Anthropic's Mike Krieger ports an internal Python tool to 165,000 lines of TypeScript over a weekend with hundreds of Claude Code agents",
      "url": "https://www.shippedbyagents.dev/stories/anthropic-mike-krieger-python-to-typescript-165k-weekend-claude-code/",
      "source_url": "https://claude.com/resources/articles/ai-code-migration",
      "source_name": "Anthropic (claude.com AI code migration article)",
      "source_type": "vendor",
      "date": "2026-07-16",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Fable 5",
        "Claude Opus 4.8",
        "Claude Sonnet"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Output",
          "value": "165,000 lines of TypeScript over a weekend"
        },
        {
          "label": "Tokens (main portion)",
          "value": "27 million"
        },
        {
          "label": "Compile time",
          "value": "~8 minutes per platform → ~2 seconds"
        },
        {
          "label": "Startup",
          "value": "6× faster binary start"
        },
        {
          "label": "Parity harness",
          "value": "7 real-world scenarios diffed against Python"
        },
        {
          "label": "Autonomous test nights",
          "value": "4 nights of Claude-designed e2e fix loops"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Compile dropped from about eight minutes per platform to about two seconds",
          "metric": "~8 min/platform → ~2 s",
          "quote": "After the port, the same compile now takes about two seconds, the binary starts 6x faster"
        },
        {
          "claim": "Binary starts 6× faster after the TypeScript port",
          "metric": "6x faster startup",
          "quote": "the binary starts 6x faster, and the team was able to retire a separate deployment pipeline"
        }
      ],
      "techniques": [
        "parity harness against original",
        "eight phase gates",
        "adversarial review rounds",
        "discard and re-run full migration",
        "smaller models for fan-out implementers",
        "overnight autonomous e2e fix loop"
      ],
      "caveats": "Vendor-published Anthropic article about Anthropic’s own tools and staff. Exact calendar date of the weekend and the tool’s product name are not stated.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-07-16 from the page's article:published_time and JSON-LD datePublished."
    },
    {
      "slug": "seif-ibrahim-11-flutter-apps-two-months-claude-code-codex-40-dollars",
      "title": "Engineer ships 11 Flutter apps and games to both app stores in two months with Claude Code and Codex, on $40 a month",
      "url": "https://www.shippedbyagents.dev/stories/seif-ibrahim-11-flutter-apps-two-months-claude-code-codex-40-dollars/",
      "source_url": "https://iseif.dev/2026/07/12/vibe-coding-for-engineers-11-apps-in-two-months/",
      "source_name": "Seif Ibrahim's blog",
      "source_type": "builder",
      "date": "2026-07-12",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [
        "Claude Sonnet 4.6",
        "GPT-5.5"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Apps shipped",
          "value": "11 on iOS and Android"
        },
        {
          "label": "Time",
          "value": "About two months (late April to end of June 2026)"
        },
        {
          "label": "AI tooling cost",
          "value": "$40 a month (two $20 plans)"
        },
        {
          "label": "Store rejections",
          "value": "0 across 11 submissions"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "11 apps designed, built and shipped to both app stores in about two months on two $20 plans",
          "metric": "11 apps; $40/month",
          "quote": "Two $20 agents were enough to design, build, test, and ship 11 apps; I never needed a $100 or $200 plan."
        }
      ],
      "techniques": [
        "one agent reviews the other's work",
        "moving to codex when claude limits hit",
        "same prompts across agents as a bake-off",
        "/clear between tasks",
        "several apps in closed testing at once"
      ],
      "caveats": "Self-reported. The apps are small, offline games and utilities. No line counts or revenue figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "matt-hodges-personal-print-magazine-pipeline-claude-code-codex",
      "title": "Matt Hodges builds a pipeline that turns his saved articles into a printed magazine, with Claude Code, Claude Design and Codex",
      "url": "https://www.shippedbyagents.dev/stories/matt-hodges-personal-print-magazine-pipeline-claude-code-codex/",
      "source_url": "https://matthodges.com/posts/2026-07-12-personal-magazine-claude-codex/",
      "source_name": "Matt Hodges (personal blog)",
      "source_type": "builder",
      "date": "2026-07-12",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time to first printed issue",
          "value": "About two weeks"
        },
        {
          "label": "Issue 001",
          "value": "Six saved articles"
        }
      ],
      "duration_days": 14,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A working pipeline and a printed personal magazine in the mailbox about two weeks later",
          "metric": "~2 weeks",
          "quote": "Two weeks later, my own personal magazine landed in my mailbox"
        }
      ],
      "techniques": [
        "vault snapshotted into git",
        "python pipeline",
        "jinja2 and weasyprint",
        "dropped an early approach that didn't work"
      ],
      "caveats": "Personal project with one user. Few hard numbers. Claude Design is not a coding tool in our list, so it isn't tagged.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "nik-mcfly-claude-plans-codex-builds-react-spa-to-nextjs-production-migration-36-hours",
      "title": "Claude plans and reviews, Codex builds: a production music site migrated from a React SPA to Next.js in 36 hours",
      "url": "https://www.shippedbyagents.dev/stories/nik-mcfly-claude-plans-codex-builds-react-spa-to-nextjs-production-migration-36-hours/",
      "source_url": "https://nikmcfly.com/blog/claude-managed-codex-coded",
      "source_name": "nikmcfly.com (Nik McFly)",
      "source_type": "builder",
      "date": "2026-07-12",
      "tools": [
        "Claude",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Migration time",
          "value": "36 hours"
        },
        {
          "label": "Codex output",
          "value": "27 commits, 7,803 lines of source code"
        },
        {
          "label": "Routes with unique metadata in initial HTML",
          "value": "0 → 45"
        },
        {
          "label": "Sitemap URLs resolving",
          "value": "9 of 15 → 45 of 45"
        },
        {
          "label": "Automated assertions per release",
          "value": "0 → 1,505"
        },
        {
          "label": "Production releases / rollbacks",
          "value": "3 / 0"
        },
        {
          "label": "Largest initial JS download",
          "value": "~1.3 MB → 658 KB"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A production site migrated from a React SPA to Next.js in 36 hours",
          "metric": "36 hours"
        },
        {
          "claim": "Routes with unique metadata in the initial HTML went from 0 to 45",
          "metric": "0 → 45"
        },
        {
          "claim": "Three production releases with no rollbacks",
          "metric": "0 rollbacks"
        },
        {
          "claim": "The owner wrote no application code",
          "quote": "I did not write application code."
        }
      ],
      "techniques": [
        "one agent implements, another reviews",
        "release contract with definition of done",
        "single typed route registry as source of truth",
        "committed evidence for every claim",
        "pixel-diff visual comparisons"
      ],
      "caveats": "Self-reported by the site owner on his own blog. A single, fairly small site. The post does not say which Claude product (Claude Code or chat) was used. A staging Docker build filled the server disk and briefly caused production API errors during the migration. The author notes the performance numbers are for this migration only.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "leo-capybara-delivery-vibe-jam-2026-winner-claude-code-15-days-27000-lines",
      "title": "iOS engineer builds Vibe Jam 2026 winner \"A Game About Capybaras Delivering Food\" in about 15 days with Claude Code: 188 commits, 27,000 lines, all code by AI",
      "url": "https://www.shippedbyagents.dev/stories/leo-capybara-delivery-vibe-jam-2026-winner-claude-code-15-days-27000-lines/",
      "source_url": "https://leocoout.medium.com/how-i-made-25k-in-15-days-with-a-game-built-entirely-by-claude-code-94b3b817f0ce",
      "source_name": "Leo (Medium)",
      "source_type": "builder",
      "date": "2026-07-09",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 4.7"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Around 15 days"
        },
        {
          "label": "Code",
          "value": "27,000 lines, 188 commits, 100% written by AI"
        },
        {
          "label": "Parallel sessions",
          "value": "2–3 Claude Code sessions at once"
        },
        {
          "label": "Result",
          "value": "Won Vibe Jam 2026 ($25,000 prize)"
        }
      ],
      "duration_days": 15,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Complete multiplayer game with in-game editors built in around 15 days, all code by AI",
          "metric": "~15 days, 27,000 lines"
        },
        {
          "claim": "Asking Claude to build editors and tools was the biggest multiplier",
          "quote": "Tooling compounds"
        }
      ],
      "techniques": [
        "plan mode before each feature",
        "rules files for constraints and conventions",
        "2-3 parallel sessions on separate code areas",
        "fresh session per feature",
        "agent builds custom dev tools"
      ],
      "caveats": "Self-reported by the builder. The game is a jam entry with about 5–10 minutes of gameplay and he does not plan to ship it on Steam. Art, 3D models, music and sound came from other AI tools (GPT Images, Grok, Tripo3D, Suno, ElevenLabs). His development cost breakdown is only in an image, so no cost is recorded here. Stack details beyond Three.js and a Cloudflare WebSocket for multiplayer are not given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "bun-zig-to-rust-rewrite-claude-code-11-days-165000-dollars",
      "title": "Bun's 535,000-line Zig codebase is ported to Rust in 11 days by up to 64 parallel Claude Code agents",
      "url": "https://www.shippedbyagents.dev/stories/bun-zig-to-rust-rewrite-claude-code-11-days-165000-dollars/",
      "source_url": "https://bun.sh/blog/bun-in-rust",
      "source_name": "Bun blog (Jarred Sumner)",
      "source_type": "vendor",
      "date": "2026-07-08",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Source size",
          "value": "535,496 lines of Zig (excluding comments), 1,448 .zig files"
        },
        {
          "label": "Time",
          "value": "11 days (May 3 to merged May 14, 2026)"
        },
        {
          "label": "Commits",
          "value": "6,778"
        },
        {
          "label": "Parallel agents",
          "value": "About 64 at peak (4 workflows x 16)"
        },
        {
          "label": "Token cost",
          "value": "Around $165,000 at API pricing (5.9B uncached input, 690M output, 72B cached input tokens)"
        },
        {
          "label": "Tests",
          "value": "100% of suite passing in CI on all platforms, 0 skipped or deleted (60,624 tests on Debian x64)"
        },
        {
          "label": "Estimate without AI",
          "value": "3 engineers about a year (author's estimate)"
        },
        {
          "label": "Bugs fixed vs. Zig version",
          "value": "128 bugs that reproduce in v1.3.14 fixed in v1.4.0"
        }
      ],
      "duration_days": 11,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "One engineer went from start to 100% of the test suite passing on all platforms in 11 days; he thinks by hand it would have taken 3 engineers about a year",
          "metric": "11 days vs ~1 year for 3 engineers",
          "quote": "By hand, I think this would've taken 3 engineers with full context on the codebase about a year"
        },
        {
          "claim": "Binary size cut by 3.8 MB on Windows, 5.5 MB on macOS and 6.8 MB on Linux",
          "metric": "-3.8 to -6.8 MB",
          "quote": "The initial changes in the Rust rewrite reduced binary size by 3.8 MB on Windows, 5.5 MB on macOS, and 6.8 MB on Linux."
        },
        {
          "claim": "Memory after 2,000 builds down from 6,745 MB to 609 MB in Bun v1.4.0",
          "metric": "6,745 → 609 MB"
        },
        {
          "claim": "Startup 10% faster on Linux",
          "metric": "10%",
          "quote": "Startup got 10% faster on Linux"
        }
      ],
      "techniques": [
        "mechanical port keeping the architecture",
        "test suite as a conformance suite",
        "2 adversarial review agents",
        "dynamic workflows",
        "about 64 parallel agents"
      ],
      "caveats": "Bun was acquired by Anthropic in December 2025 and the author works at Anthropic, so this is close to vendor-published. Used a pre-release Anthropic model (Claude Fable 5); tokens were not paid for. The time-without-AI figure is the author's own estimate. Output is a faithful port, not idiomatic Rust (about 4% of code sits inside unsafe blocks).",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cmu-stanford-enterprise-2x-mandate-802-developers-ai-doubles-merged-prs",
      "title": "CMU and Stanford study of a company's \"2x\" AI mandate: merged PRs per engineer doubled, and automated review overtook human review",
      "url": "https://www.shippedbyagents.dev/stories/cmu-stanford-enterprise-2x-mandate-802-developers-ai-doubles-merged-prs/",
      "source_url": "https://arxiv.org/html/2607.01904",
      "source_name": "arXiv 2607.01904",
      "source_type": "independent",
      "date": "2026-07-02",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Sample",
          "value": "802 developers, 196,212 pull requests (Jan 2024 to Apr 2026)"
        },
        {
          "label": "Throughput",
          "value": "2.09x pre-mandate merged PRs per engineer by April 2026"
        },
        {
          "label": "Review load",
          "value": "Per-reviewer load roughly doubled"
        },
        {
          "label": "Quality signals",
          "value": "Merge and revert rates held steady"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Merged PRs per engineer reached about 2x the pre-mandate baseline, but only ten months after the mandate",
          "metric": "2.09x",
          "quote": "Per-capita PR throughput approached 2× only in April 2026, ten months after the announcement"
        },
        {
          "claim": "PR throughput rose about 12% when a developer started using AI, then kept growing with use",
          "metric": "+12%",
          "quote": "PR throughput rises by about 12% when a developer starts using AI and then keeps growing with accumulated use"
        }
      ],
      "techniques": [
        "staggered difference-in-differences",
        "company-wide ai mandate",
        "merged prs per engineer as the target metric"
      ],
      "caveats": "One anonymous company. Adoption was not randomly assigned, so the authors avoid exact causal attribution. The abstract does not name the AI tools.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-07-02 from the arXiv submission date on abs/2607.01904."
    },
    {
      "slug": "microsoft-claude-code-copilot-cli-rollout-24-percent-more-merged-prs",
      "title": "Microsoft study of its early-2026 Claude Code and Copilot CLI rollout: adopters merged about 24% more pull requests",
      "url": "https://www.shippedbyagents.dev/stories/microsoft-claude-code-copilot-cli-rollout-24-percent-more-merged-prs/",
      "source_url": "https://arxiv.org/html/2607.01418",
      "source_name": "arXiv 2607.01418",
      "source_type": "vendor",
      "date": "2026-07-01",
      "tools": [
        "Claude Code",
        "GitHub Copilot"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Population",
          "value": "Tens of thousands of Microsoft engineers"
        },
        {
          "label": "Merged PR lift",
          "value": "+24.0% PRs per engineer per day (95% CI +14.5% to +33.7%)"
        },
        {
          "label": "Lift by tool, any-use weeks",
          "value": "Copilot CLI +24.9%, Claude Code +11.4%"
        },
        {
          "label": "Window",
          "value": "Four months after rollout"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Adopters merged roughly 24% more pull requests than they would have otherwise",
          "metric": "+24%",
          "quote": "adopters merged roughly 24% more pull requests than they would"
        }
      ],
      "techniques": [
        "adoption spreads through peers",
        "visible peer use as rollout strategy"
      ],
      "caveats": "Written by Microsoft researchers, and Microsoft owns GitHub Copilot, so labelled vendor. The authors note a merged PR is not the same as the value it delivers.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-07-01 from the arXiv submission date on abs/2607.01418."
    },
    {
      "slug": "gusto-cofounder-five-person-team-claude-code-10-weeks-tier-one-launch",
      "title": "Gusto's CTO and a five-person team built Gusto Cofounder, a new AI product, with Claude Code from zero code to tier-one launch in 10 weeks",
      "url": "https://www.shippedbyagents.dev/stories/gusto-cofounder-five-person-team-claude-code-10-weeks-tier-one-launch/",
      "source_url": "https://www.lennysnewsletter.com/p/no-figma-no-jira-no-docs-how-gusto",
      "source_name": "How I AI podcast (Claire Vo) with Eddie Kim, Lenny's Newsletter",
      "source_type": "builder",
      "date": "2026-06-29",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Team",
          "value": "Four engineers (including the CTO) and one designer"
        },
        {
          "label": "Time",
          "value": "10 weeks from zero code to tier-one launch"
        },
        {
          "label": "Designer output",
          "value": "94th percentile for shipping code"
        }
      ],
      "duration_days": 70,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Five people took a new product from zero code to tier-one launch in 10 weeks",
          "metric": "10 weeks",
          "quote": "from zero code to a tier-one launch in 10 weeks"
        },
        {
          "claim": "A designer with no engineering background reached the 94th percentile for shipping code",
          "metric": "94th percentile"
        }
      ],
      "techniques": [
        "prototype in code before deciding",
        "write a failing eval before the fix",
        "no specs or tickets",
        "permanent shared video room"
      ],
      "caveats": "A podcast interview with the CTO who led the work; figures come from the episode page and transcript. No figures on lines of code, cost, or customer results. The 94th-percentile figure is not defined on the page.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "atlassian-spec-driven-graphql-fetcher-migration-two-quarters-one-week",
      "title": "Atlassian's Jira team migrates 100+ GraphQL fetchers off a legacy gateway in a week, about two quarters of work, using agents driven by Jira specs",
      "url": "https://www.shippedbyagents.dev/stories/atlassian-spec-driven-graphql-fetcher-migration-two-quarters-one-week/",
      "source_url": "https://www.atlassian.com/blog/development/spec-driven-ai-migration",
      "source_name": "Inside Atlassian blog (Raymond Su, Reshad Contractor)",
      "source_type": "builder",
      "date": "2026-06-25",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Scope",
          "value": "100+ GraphQL fetchers"
        },
        {
          "label": "Team",
          "value": "5 developers"
        },
        {
          "label": "Time",
          "value": "\"around 2 quarters worth of work in a week\""
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Five developers did around two quarters of work in a week, moving 100+ GraphQL fetchers",
          "metric": "~2 quarters in 1 week",
          "quote": "We managed to do this with 5 developers completing around 2 quarters worth of work in a week"
        }
      ],
      "techniques": [
        "ai writes a behavioural spec from existing code",
        "engineers review each spec",
        "jira tickets as specs",
        "parallel batches",
        "front-load human judgment into the spec"
      ],
      "caveats": "Self-reported on Atlassian's own blog, and the post doubles as a pitch for Jira as agent infrastructure. It names no agent, model or vendor, and gives no line counts, cost, defect or rollback numbers. The two-quarters figure is the team's estimate of manual effort.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "flywheel-solo-staff-engineer-325k-line-platform-11-months-claude-code-2900-sessions",
      "title": "Solo staff engineer builds a 325,000-line workflow automation platform in 11 months of evenings and weekends with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/flywheel-solo-staff-engineer-325k-line-platform-11-months-claude-code-2900-sessions/",
      "source_url": "https://dev.to/traviticus/i-built-a-325k-line-enterprise-platform-in-11-months-solo-heres-how-ai-changed-whats-possible-2c2c",
      "source_name": "Travis Wilson (DEV Community)",
      "source_type": "builder",
      "date": "2026-06-25",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "325,000 lines of production code (170K Go, 145K TypeScript, 7.5K Python, 2,500 Terraform)"
        },
        {
          "label": "Time",
          "value": "11 months, solo, evenings and weekends"
        },
        {
          "label": "Commits",
          "value": "702"
        },
        {
          "label": "Tests",
          "value": "8,000+ tests, 847 test files"
        },
        {
          "label": "Claude Code usage",
          "value": "2,900 sessions across 144 active days"
        },
        {
          "label": "Cost",
          "value": "~$1,200 in Claude Max subscription (would have metered at $2,500+ on API pricing)"
        },
        {
          "label": "Churn",
          "value": "1.37 million lines inserted, 716K deleted"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "One person built a 325K-line platform with 14 integrations in 11 months",
          "metric": "325,000 lines, 11 months",
          "quote": "Production code with tests, CI/CD, infra, and 14 integrations. One person. Eleven months."
        },
        {
          "claim": "Claude Code usage cost about $1,200, versus $2,500+ at API prices",
          "metric": "~$1,200 Claude Max"
        }
      ],
      "techniques": [
        "claude sketches, i refine",
        "claude challenges its own plan",
        "tests after manual verification",
        "claude.md context files",
        "continuous refactoring"
      ],
      "caveats": "Self-reported by the builder, who is promoting his product. The post says he is shipping to real users but gives no user or revenue numbers, and 1.0 had not shipped yet. Lines of code are a size measure, not an outcome. The year of the DEV post is not shown on the page; June 2026 is inferred from the post saying \"Now (June 2026)\".",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "datadog-stream-router-kv-to-postgresql-migration-claude-cursor-45-minutes-to-1-second",
      "title": "Datadog migrates its live Stream Router from FoundationDB to PostgreSQL with Claude and Cursor; 45-minute operations now take ~1 second",
      "url": "https://www.shippedbyagents.dev/stories/datadog-stream-router-kv-to-postgresql-migration-claude-cursor-45-minutes-to-1-second/",
      "source_url": "https://www.datadoghq.com/blog/engineering/ai-assisted-storage-migration/",
      "source_name": "Datadog engineering blog (Arnold Wakim)",
      "source_type": "builder",
      "date": "2026-06-23",
      "tools": [
        "Claude",
        "Cursor"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Design to production",
          "value": "~3 months (Dec 2025 – Feb 2026); proof of concept in 4 weeks"
        },
        {
          "label": "Slowest operations",
          "value": "Estimated 45 minutes → ~1 second"
        },
        {
          "label": "Routing dataset size",
          "value": "40x smaller in PostgreSQL"
        },
        {
          "label": "Latency",
          "value": "Hundreds of ms → a few ms"
        },
        {
          "label": "Database cost",
          "value": "Down 90%"
        },
        {
          "label": "Production incidents during deployment",
          "value": "None"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Operations estimated at 45 minutes now complete in about 1 second",
          "metric": "45 minutes → ~1 second"
        },
        {
          "claim": "Database costs fell 90% after retiring FoundationDB",
          "metric": "-90% database cost"
        },
        {
          "claim": "Deployment to production proceeded without incident",
          "quote": "Deployment to production proceeded without incident."
        },
        {
          "claim": "The team says it would not have finished in time without AI tools",
          "quote": "We would not have completed the implementation in the time frame we had without AI tools."
        }
      ],
      "techniques": [
        "ai writes intent docs before refactoring",
        "stub, failing test, iterate per method",
        "one method per prompt",
        "blue/green validation against live traffic",
        "human-written optimised sql as examples"
      ],
      "caveats": "Self-reported by the Datadog team that did the work. The AI tools were not autonomous: humans designed the schema, reviewed every generated skeleton, and wrote the optimised SQL (batching, UNNEST, CTEs), which the models then copied. Most of the performance and cost gains come from the new relational design, not from the AI tools; the post credits AI with making the implementation possible in the time frame. No count of methods, PRs or lines is given, and token spend is described only as significant.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "mercari-ios-ds3-to-ds4-design-system-migration-22-claude-code-sessions-140-prs",
      "title": "A Mercari iOS engineer runs 22 Claude Code sessions in git worktrees to migrate screens to a new design system, opening about 140 PRs",
      "url": "https://www.shippedbyagents.dev/stories/mercari-ios-ds3-to-ds4-design-system-migration-22-claude-code-sessions-140-prs/",
      "source_url": "https://engineering.mercari.com/en/blog/entry/20260619-scaling-myself-how-i-run-22-claude-code-sessions-for-ds4-migration/",
      "source_name": "Mercari Engineering (cyan)",
      "source_type": "builder",
      "date": "2026-06-22",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs",
          "value": "Around 140 over several months; up to 18 open at once"
        },
        {
          "label": "Sessions",
          "value": "22 Claude Code sessions (18 worktree, 4 planning/review)"
        },
        {
          "label": "Screens double-checked",
          "value": "21 screens; 3 had hidden DS3 dependencies"
        },
        {
          "label": "Check speed",
          "value": "About a minute per screen for modify-build-restore cycles"
        },
        {
          "label": "Setup time",
          "value": "About two months of iteration"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Sustained around 140 migration PRs, with up to 18 open at the same time",
          "metric": "~140 PRs",
          "quote": "I have created around 140 Pull Requests (PRs), and at one point I had 18 of them open at the same time."
        },
        {
          "claim": "Simulated cleanup check found 3 of 21 screens with hidden dependencies, at about a minute per screen",
          "metric": "3 of 21 screens",
          "quote": "running 21 sequential modify-build-restore cycles took about a minute per screen"
        }
      ],
      "techniques": [
        "git worktree per session",
        "slash commands encode migration phases",
        "shared bazel cache with build lock",
        "one session owns each stacked pr chain",
        "separate planning and execution sessions"
      ],
      "caveats": "Self-reported by one engineer. No figures on total time saved, defects or how long the migration would have taken by hand. The 'several months' span and PR count are approximate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "forter-jenkins-to-github-actions-500-repos-three-months-claude-code",
      "title": "Forter migrates about 90% of 500 repos from Jenkins to GitHub Actions in under three months with a local Claude Code migration agent",
      "url": "https://www.shippedbyagents.dev/stories/forter-jenkins-to-github-actions-500-repos-three-months-claude-code/",
      "source_url": "https://blog.forter.dev/genai-ci-migrations-part-1-local-agentic-flows/",
      "source_name": "Forter Tech Blog (Dor Almog)",
      "source_type": "builder",
      "date": "2026-06-16",
      "tools": [
        "Claude Code",
        "GitHub Copilot"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Repos migrated",
          "value": "About 91% of 500, including 100% of non-custom repos"
        },
        {
          "label": "Calendar time",
          "value": "Just under three months, against an estimate of about a year"
        },
        {
          "label": "Effort",
          "value": "About 10 weeks, never more than 3 engineers at a time"
        },
        {
          "label": "Original estimate",
          "value": "3.4 engineering years"
        },
        {
          "label": "Claimed speed-up",
          "value": "13x"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 91% of 500 repos moved to GitHub Actions in just under three months, against an estimate of roughly a year",
          "metric": "~3 months vs ~1 year",
          "quote": "we managed to fully migrate ~91% of all repositories, including 100% of our goal repos, in just under three months"
        },
        {
          "claim": "Forter says its framework sped up the migration 13 times over the 3.4 engineering-year estimate",
          "metric": "13x",
          "quote": "sped up a full-scale CI migration 13 times over"
        }
      ],
      "techniques": [
        "claude code skill",
        "four-phase analysis agent workflow",
        "migration paths in the skill",
        "context docs on old and new ci"
      ],
      "caveats": "Company writing about its own work. Part 1 of a 3-part series; the three-month clock starts once the workflow was stable, not from the first attempt.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "fesom2-ocean-model-fortran-to-c-to-kokkos-claude-code-74000-lines-weeks",
      "title": "Alfred Wegener Institute ports the FESOM2 ocean model (about 74,000 lines of Fortran) to C and then GPU-ready C++/Kokkos with Claude Code in a matter of weeks",
      "url": "https://www.shippedbyagents.dev/stories/fesom2-ocean-model-fortran-to-c-to-kokkos-claude-code-74000-lines-weeks/",
      "source_url": "https://arxiv.org/abs/2606.11356",
      "source_name": "arXiv (Koldunov, Cheedela, Danilov, Sidorenko, Beyer, Jung)",
      "source_type": "builder",
      "date": "2026-06-09",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 4.7"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Fortran ported",
          "value": "73,991 lines in 109 files (core ocean–sea-ice configuration)"
        },
        {
          "label": "C port",
          "value": "20,010 lines, 73 files, 10 days with commits"
        },
        {
          "label": "C++/Kokkos port",
          "value": "30,663 lines, 78 files, 7 days with commits"
        },
        {
          "label": "Fidelity, C vs Fortran over five years",
          "value": "SST RMS difference 0.006 °C, SSS 0.002 PSU"
        },
        {
          "label": "GPU vs CPU node",
          "value": "1.6–3.7x faster on meshes up to 7.4 million surface vertices"
        },
        {
          "label": "Throughput at 32 nodes",
          "value": "1.3 SYPD (A100) to 3.5 SYPD (GH200)"
        },
        {
          "label": "Bring-up on other GPU systems",
          "value": "Less than a day each"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Full Fortran ocean model moved to another language and onto GPUs while preserving its physics in a matter of weeks",
          "quote": "while preserving its physics in a matter of weeks"
        },
        {
          "claim": "A single A100 GPU node runs 1.6–3.7x faster than a CPU node on eddy-rich meshes",
          "metric": "1.6–3.7x"
        },
        {
          "claim": "Reaches the 1–2 simulated-years-per-day production band on every machine tested",
          "metric": "1.3–3.5 SYPD at 32 nodes"
        },
        {
          "claim": "Each new accelerator system took less than a day to compile and run",
          "metric": "< 1 day per system"
        }
      ],
      "techniques": [
        "literal line-by-line translation",
        "two-stage port via c reference",
        "validation ladder per stage",
        "plan, lessons-learned and handoff files",
        "short narrowly scoped sessions"
      ],
      "caveats": "Written by the scientists who directed the port. The ported configuration covers the core ocean–sea-ice model (the full distribution is about 269,000 lines) and still lacks restart I/O, other vertical-coordinate variants and coupled-climate physics needed to replace the Fortran in production. Days with commits are not total effort, and no cost or human hours are given. An earlier attempt by the first author had failed. The output system was redesigned rather than ported.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "atolio-svelte-4-to-5-thousands-of-files-claude-code-six-hours-one-bug",
      "title": "Atolio migrates thousands of files from Svelte 4 to 5 with Claude Code in a six-hour run, QA finds one bug",
      "url": "https://www.shippedbyagents.dev/stories/atolio-svelte-4-to-5-thousands-of-files-claude-code-six-hours-one-bug/",
      "source_url": "https://www.atolio.com/blog/svelte-4-to-5-migration-with-claude-code",
      "source_name": "Atolio blog (Pablo Arellano)",
      "source_type": "builder",
      "date": "2026-06-04",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Files migrated",
          "value": "Thousands"
        },
        {
          "label": "Agent run",
          "value": "About 6 hours in auto mode"
        },
        {
          "label": "Follow-up QA",
          "value": "About 2 days, 2 developers"
        },
        {
          "label": "Bugs found in QA",
          "value": "1"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A migration of thousands of files came down to a single bug in QA",
          "metric": "1 bug",
          "quote": "Once it was in testing, the team found only one bug during QA."
        }
      ],
      "techniques": [
        "months of human investigation first",
        "detailed prompt from gathered findings",
        "plan mode review",
        "auto mode run",
        "code-only access"
      ],
      "caveats": "Written by the developer who ran it. No exact file count, line count or cost. Months of human preparation came before the six-hour run.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "bloomberg-pomona-agent-skills-small-code-quality-prs-82-percent-merged",
      "title": "Bloomberg's Pomona agent opens one tiny code-quality PR a day, and 32 of 39 were merged with a median close time of just over two hours",
      "url": "https://www.shippedbyagents.dev/stories/bloomberg-pomona-agent-skills-small-code-quality-prs-82-percent-merged/",
      "source_url": "https://arxiv.org/abs/2606.06752",
      "source_name": "arXiv 2606.06752 (Williams et al., Bloomberg and UCL)",
      "source_type": "builder",
      "date": "2026-06-04",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs",
          "value": "39 over 55 workdays; 32 merged (82.1%)"
        },
        {
          "label": "Median time to close",
          "value": "2 hours 14 minutes"
        },
        {
          "label": "Merged as-is",
          "value": "28 of 32 with no extra commits; 25 with no interaction beyond mandatory review"
        },
        {
          "label": "Engineer survey",
          "value": "10 of 12 senior engineers wanted to try it"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 32,
      "claimed_benefits": [
        {
          "claim": "82.1% of PRs merged, 32 of 39",
          "metric": "32/39 (82.1%)"
        },
        {
          "claim": "Median time-to-close of just over two hours",
          "metric": "2h 14m"
        },
        {
          "claim": "10 of 12 surveyed senior engineers wanted to adopt it",
          "metric": "10/12"
        }
      ],
      "techniques": [
        "agent skills",
        "prioritised backlog in repo root",
        "about 10 lines of diff per pr",
        "one pr per day",
        "human review required before merge"
      ],
      "caveats": "Industrial experience paper by Bloomberg engineers and UCL researchers about their own tool, one team, small sample. The paper says it cannot reveal which AI services or models were used, and data and code are not released.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "uber-ai-coding-budget-cap",
      "title": "Uber caps AI coding spend at $1,500 per employee per tool after burning its annual AI budget in 4 months",
      "url": "https://www.shippedbyagents.dev/stories/uber-ai-coding-budget-cap/",
      "source_url": "https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/",
      "source_name": "TechCrunch (reporting Bloomberg and The Information)",
      "source_type": "independent",
      "date": "2026-06-02",
      "tools": [
        "Claude Code",
        "Cursor"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Monthly cap",
          "value": "$1,500 per employee, per tool"
        },
        {
          "label": "Time to exhaust annual AI budget",
          "value": "4 months"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "monthly spending cap per employee per tool",
        "internal usage leaderboards"
      ],
      "caveats": "Press report based on Bloomberg and The Information. Uber's COO said it is 'very hard to draw a line' between AI usage and new consumer features.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "sixt-ios-xctest-to-swift-testing-custom-claude-agent-780-files-108914-lines-106-dollars",
      "title": "Sixt's custom Claude agent migrates 780 iOS test files (~7,000 tests, 108,914 lines) from XCTest to Swift Testing for about $106",
      "url": "https://www.shippedbyagents.dev/stories/sixt-ios-xctest-to-swift-testing-custom-claude-agent-780-files-108914-lines-106-dollars/",
      "source_url": "https://medium.com/@ibrahimoktay/at-sixt-we-migrated-7000-unit-tests-with-ai-the-cost-about-106-13416c6f7a94",
      "source_name": "Medium (İbrahim Oktay, Sixt staff engineer)",
      "source_type": "builder",
      "date": "2026-06-02",
      "tools": [
        "Claude Agent SDK"
      ],
      "models": [
        "Claude Opus 4.5"
      ],
      "work_type": "Test migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Files processed",
          "value": "780 across 50 modules"
        },
        {
          "label": "Unit tests migrated",
          "value": "~7,000"
        },
        {
          "label": "Lines changed",
          "value": "108,914"
        },
        {
          "label": "Success rate",
          "value": "97.3% (759 of 780; 21 failed)"
        },
        {
          "label": "Total compute",
          "value": "~3.8 hours over 61 runs"
        },
        {
          "label": "Total AI cost",
          "value": "~$106 (~$0.14 per file)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "780 test files across 50 modules migrated with a 97.3% success rate",
          "metric": "97.3% of 780 files"
        },
        {
          "claim": "108,914 lines changed for about $106 in AI cost",
          "metric": "~$106",
          "quote": "A single engineer day costs more than the entire migration did."
        },
        {
          "claim": "About 3.8 hours of agent compute versus an estimated months of manual work",
          "metric": "~3.8 hours compute"
        }
      ],
      "techniques": [
        "cheap model pass, escalate to research agent",
        "swiftc -parse validation with error feedback",
        "learned-pattern memory file",
        "fixer agent runs real test suite",
        "checkpoint and resume"
      ],
      "caveats": "Written by the engineer who built the agent, on a personal Medium blog, with recruiting-style framing. Figures come from the agent's own cumulative stats. The author says the 21 failures were nearly all network errors to the LLM gateway, not migration mistakes. The manual-effort comparison (\"months\", 5–10 files a day per senior engineer) is the author's estimate. The migration is not fully completed or merged yet. Sixt used its internal LLM gateway; the post names the Anthropic SDK for bulk transforms and the Claude Agent SDK for the research and fixer agents.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "shopify-river-slack-agent-3536-merged-prs-30-days-one-in-eight",
      "title": "Shopify's Slack-native agent River co-authors one in eight merged pull requests, with 3,536 merged in 30 days",
      "url": "https://www.shippedbyagents.dev/stories/shopify-river-slack-agent-3536-merged-prs-30-days-one-in-eight/",
      "source_url": "https://shopify.engineering/under-the-river",
      "source_name": "Shopify Engineering blog (Javier Moreno, co-authored by River)",
      "source_type": "builder",
      "date": "2026-05-28",
      "tools": [],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of merged PRs",
          "value": "\"one in eight merged pull requests across Shopify is coauthored by it\""
        },
        {
          "label": "Merged PRs",
          "value": "3,536 River-coauthored PRs in a recent 30-day period"
        },
        {
          "label": "Sessions",
          "value": "59,918 sessions in 5,170 Slack channels in that 30 days, touching the work of 7,000+ people"
        },
        {
          "label": "Session size",
          "value": "median 19 minutes and 50 tool calls"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "One in eight merged pull requests across Shopify is coauthored by River, with 3,536 merged in a recent 30-day period",
          "metric": "1 in 8 PRs; 3,536 in 30 days",
          "quote": "3,536 River-coauthored pull requests merged in that window"
        }
      ],
      "techniques": [
        "agent called in public slack channels",
        "no direct messages",
        "public transcripts mined to improve skills"
      ],
      "caveats": "Shopify's own engineering blog, co-authored by the agent itself; numbers come from River's own session table. It names no model or vendor, and gives no cost, review effort, or quality numbers for River's PRs. \"Co-authored\" doesn't say how much of each PR River wrote.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dropbox-nova-internal-coding-agent-platform-1-in-12-pull-requests-flaky-tests-migrations",
      "title": "Dropbox’s internal coding agent platform Nova accounts for roughly 1 in 12 pull requests, and runs flaky-test fixes and migrations",
      "url": "https://www.shippedbyagents.dev/stories/dropbox-nova-internal-coding-agent-platform-1-in-12-pull-requests-flaky-tests-migrations/",
      "source_url": "https://dropbox.tech/culture/beyond-code-generation-rethinking-engineering-productivity-in-the-age-of-ai-agents",
      "source_name": "Dropbox Tech blog (Kazuaki Okumura, DX Annual 2026)",
      "source_type": "builder",
      "date": "2026-05-28",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of pull requests",
          "value": "Roughly 1 in 12 pull requests at Dropbox"
        },
        {
          "label": "Flaky test fix validation",
          "value": "Test run 100 or more times in CI; up to five fix attempts"
        },
        {
          "label": "Earlier Goose-based migrator",
          "value": "Used across thousands of migration entries"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Nova accounts for roughly 1 in 12 pull requests at Dropbox",
          "metric": "1 in 12 PRs",
          "quote": "accounting for roughly 1 in 12 pull requests at Dropbox today"
        }
      ],
      "techniques": [
        "validation commands fed back to the agent",
        "flaky test fix rerun 100+ times in ci",
        "one branch per session, publishing outside the agent",
        "localized agents.md files",
        "deterministic ci triggers, agent called back on failure"
      ],
      "caveats": "Written by Dropbox engineers. The 1 in 12 figure comes from the DX Annual write-up and has no date range or merge rate attached; the Nova introduction post gives no output numbers. Dropbox does not name the coding agents Nova runs today, only that it supports several; the earlier migrator was Goose-based.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "salesforce-claude-code-company-wide-33-api-endpoints-13-days",
      "title": "Salesforce moves its whole engineering org to Claude Code and says a 231-person-day API migration took 13 days",
      "url": "https://www.shippedbyagents.dev/stories/salesforce-claude-code-company-wide-33-api-endpoints-13-days/",
      "source_url": "https://www.salesforce.com/news/stories/how-engineering-became-agentic/",
      "source_name": "Salesforce News (Srinivas Tallapragada, President and Chief Engineering and Customer Success Officer)",
      "source_type": "builder",
      "date": "2026-05-27",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "API migration",
          "value": "33 endpoints in 13 days vs ~231 person-days estimated (18x)"
        },
        {
          "label": "Merged PRs per developer",
          "value": "Up 79% (April 2026 vs April 2025)"
        },
        {
          "label": "Completed work items per developer",
          "value": "Up 50.8%"
        },
        {
          "label": "Incidents",
          "value": "Down 5%"
        },
        {
          "label": "Migration PRs",
          "value": "Five, the largest covering 21 endpoints with full test coverage"
        }
      ],
      "duration_days": 13,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "33 API endpoints migrated in 13 days, against roughly 231 person-days the traditional way",
          "metric": "13 days vs ~231 person-days",
          "quote": "They did it in 13 days."
        },
        {
          "claim": "Per developer, completed work items up 50.8% and merged PRs up 79%, April 2026 against April 2025",
          "metric": "+50.8% items; +79% PRs",
          "quote": "PRs merged per developer are up 79%."
        },
        {
          "claim": "Total incidents down 5% even with more PRs",
          "metric": "-5% incidents",
          "quote": "Even with the increase in PRs, total incidents dropped by 5%."
        }
      ],
      "techniques": [
        "claude code as the main agent",
        "unlimited tokens for every developer",
        "ml-based effective output score"
      ],
      "caveats": "Salesforce writing about its own engineering org; The Decoder notes none of the numbers can be independently verified. The 231 person-days is the team's own estimate (seven per API). The 13-day migration came out as five PRs, the largest covering 21 endpoints with 100% test coverage. No cost or token figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stripe-autojdk-java-upgrades-minions-160-prs-93-percent-jdk17",
      "title": "Stripe builds continuous Java upgrades with Minions fixing CI failures, reaching 93%+ of services on JDK 17 within a week of rollout",
      "url": "https://www.shippedbyagents.dev/stories/stripe-autojdk-java-upgrades-minions-160-prs-93-percent-jdk17/",
      "source_url": "https://stripe.dev/blog/modern-java-at-stripe-language-upgrades-as-a-service",
      "source_name": "Stripe Dot Dev blog (Mahib Hosain, Cara Mecozzi)",
      "source_type": "builder",
      "date": "2026-05-27",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "JDK 17 adoption",
          "value": "93%+ of services within under a week of rollout"
        },
        {
          "label": "PRs merged",
          "value": "160+ in three months"
        },
        {
          "label": "Legacy JDK footprint",
          "value": "Down by roughly 25K targets; all Java 8 and most Java 11 services gone"
        },
        {
          "label": "Largest PRs",
          "value": "About 10K targets upgraded per PR"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "JDK 17 adoption reached 93%+ of services in under a week of rollout",
          "metric": "93%+",
          "quote": "In under 1 week of rollout progress, service-level adoption reached 93%+ on JDK 17"
        },
        {
          "claim": "160+ PRs merged and the legacy-JDK footprint cut by roughly 25K targets, with Java 8 services eliminated",
          "metric": "160+ PRs; ~25K targets",
          "quote": "Shrunk the legacy-JDK footprint, specifically JDK 8 and 11, by roughly 25K targets"
        }
      ],
      "techniques": [
        "reads the live build graph",
        "bounded, reviewable prs",
        "agents fix failing ci",
        "fresh git checkouts in sandboxes",
        "tracks blockers"
      ],
      "caveats": "Much of the work is deterministic build-graph automation; agents handle a subset of failures. The post doesn't say how many fixes came from Minions or which model they use.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "plaid-hamilton-cloud-coding-agent-20-percent-of-merges",
      "title": "Plaid’s Hamilton cloud coding agent writes thousands of PRs and accounts for about 20% of all merges",
      "url": "https://www.shippedbyagents.dev/stories/plaid-hamilton-cloud-coding-agent-20-percent-of-merges/",
      "source_url": "https://engineering.plaid.com/hamilton-our-cloud-coding-agent-7db1c1eae764",
      "source_name": "Plaid Engineering (Clay Allsopp)",
      "source_type": "builder",
      "date": "2026-05-19",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Hamilton share of merges",
          "value": "Roughly 20% of all merges"
        },
        {
          "label": "Volume",
          "value": "Thousands of PRs written non-stop"
        },
        {
          "label": "Platform uplift",
          "value": "~3× daily cloud-agent PR volume vs first CI platform"
        },
        {
          "label": "Company AI PRs (May 2026 sister post)",
          "value": "80% of PRs AI-generated (~90% of code); +50% PRs per engineer"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Hamilton accounts for roughly 20% of all merges",
          "metric": "~20% of merges",
          "quote": "Hamilton now non-stop writes thousands of PRs for our engineers and accounts for roughly 20% of all merges."
        },
        {
          "claim": "New Hamilton platform ~3×’d daily cloud agent PR volume versus the original CI-based platform",
          "metric": "~3× daily cloud PR volume",
          "quote": "Hamilton on our new platform immediately ~3x’d our daily cloud agent PR volume as compared to our original platform"
        }
      ],
      "techniques": [
        "reuse local harnesses in cloud",
        "multi-agent orchestration",
        "ci autofix agent",
        "slack and github triggers",
        "skills kick off cloud agents"
      ],
      "caveats": "Self-reported. Commercial coding harnesses are not named (tools left Unstated; Hamilton is internal). The 20% merge share and “thousands of PRs” lack absolute denominators. Sister adoption post is a secondary URL for company-wide context.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "luke-lin-baohua-stories-app-one-month-claude-code-codex-review",
      "title": "Luke Lin ships Baohua, a bilingual story-sharing app for families, in about a month with Claude Code writing the code and Codex reviewing PRs",
      "url": "https://www.shippedbyagents.dev/stories/luke-lin-baohua-stories-app-one-month-claude-code-codex-review/",
      "source_url": "https://pmdata.substack.com/p/the-solo-production-stack",
      "source_name": "The AI Craft (Luke Lin's Substack)",
      "source_type": "builder",
      "date": "2026-05-06",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time to ship",
          "value": "April 1 to April 30, 2026"
        },
        {
          "label": "Tickets",
          "value": "Roughly 80 at peak across three milestones"
        },
        {
          "label": "Codex review",
          "value": "Two to three critical issues flagged on a typical PR"
        },
        {
          "label": "Test gate",
          "value": "Roughly one in three PRs fails a test on first push"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A bilingual story-sharing app taken from a fresh session to release in a month, April 1 to April 30"
        }
      ],
      "techniques": [
        "gstack skills",
        "agent interviews you about the product",
        "codex reviews each pr",
        "test gate on push",
        "tickets across milestones"
      ],
      "caveats": "Self-reported by the builder. No user or revenue numbers. The post recommends tools by name (gstack, Venn AI).",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "google-tensorflow-to-jax-model-migration-multi-agent-system-6x-faster",
      "title": "Google migrates production YouTube models from TensorFlow to JAX with a planner, orchestrator and coder agent system, 6.4x to 8x faster than by hand",
      "url": "https://www.shippedbyagents.dev/stories/google-tensorflow-to-jax-model-migration-multi-agent-system-6x-faster/",
      "source_url": "https://cloud.google.com/blog/topics/developers-practitioners/6x-faster-migration-from-tensorflow-to-jax",
      "source_name": "Google Cloud Blog (Jamie Rogers, Parthasarathy Ranganathan)",
      "source_type": "builder",
      "date": "2026-05-06",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Speedup",
          "value": "6.4x to 8x over manual migration"
        },
        {
          "label": "Time",
          "value": "Several SWE-months reduced to a few weeks of AI-assisted generation plus expert review"
        },
        {
          "label": "Models migrated",
          "value": "Real-world YouTube models with thousands of lines of code and hundreds of layers"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Model migration 6.4x to 8x faster than doing it manually",
          "metric": "6.4x to 8x",
          "quote": "the multi-agent system achieved a 6.4x to 8x speedup over performing the migration manually"
        },
        {
          "claim": "Several SWE-months of work reduced to a few weeks plus expert review",
          "metric": "several SWE-months to a few weeks"
        }
      ],
      "techniques": [
        "planner with compiler-based static analysis",
        "migrate from leaf nodes upward",
        "test-and-fix loop",
        "playbooks of golden examples",
        "numerical equivalence check per layer"
      ],
      "caveats": "Google writing about its own system; it also sells AI products. The post does not name the coding model or say how many models were migrated. Speedup is from their evaluations; a technical paper is referenced but not linked here.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "nber-writing-code-vs-shipping-code-ai-agents-500k-developers",
      "title": "NBER study of 500,000+ GitHub developers: coding agents multiply commits and code, but releases rise only ~20%",
      "url": "https://www.shippedbyagents.dev/stories/nber-writing-code-vs-shipping-code-ai-agents-500k-developers/",
      "source_url": "https://www.nber.org/system/files/working_papers/w35275/w35275.pdf",
      "source_name": "NBER Working Paper No. 35275",
      "source_type": "independent",
      "date": "2026-05",
      "tools": [
        "GitHub Copilot",
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Developers studied",
          "value": "500,000+"
        },
        {
          "label": "Weekly commits",
          "value": "+30% autocomplete, ~+180% with sync agents, ~+240% with async agents (cumulative)"
        },
        {
          "label": "Sync agents: lines of code",
          "value": "+958%"
        },
        {
          "label": "Sync agents: pull requests",
          "value": "+86%"
        },
        {
          "label": "Sync agents: releases",
          "value": "+20%"
        },
        {
          "label": "Lines added vs deleted after sync agents",
          "value": "7.0x added, 12.2x deleted"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Weekly commits up about 30% with autocomplete, 180% with sync agents added and 240% with async agents added",
          "metric": "+240% commits",
          "quote": "Autocomplete raises weekly commits by about 30%; adding sync agents brings the cumulative effect to about 180%; adding async agents brings it to about 240%."
        },
        {
          "claim": "Sync agents raised lines of code 958% and pull requests 86%, but releases only 20%",
          "metric": "+20% releases",
          "quote": "Sync agents raise lines of code by 958% and pull requests by 86%, but releases by only 20%."
        }
      ],
      "techniques": [
        "matched event study",
        "sync agents vs async agents"
      ],
      "caveats": "Observational event study, not an RCT. Uses internal Microsoft telemetry, and NBER's disclosures say Demirer and Musolff previously held postdoc positions at Microsoft and Demirer is a paid research consultant to the company. Revised September 2026.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "sergio-azocar-claude-code-monorepo-toolchain-migration-700-files-lint-typecheck-23x-faster",
      "title": "Sergio Azócar uses Claude Code to move a 1,670-file monorepo to oxlint and tsgo, cutting lint plus typecheck from 29s to 1.25s",
      "url": "https://www.shippedbyagents.dev/stories/sergio-azocar-claude-code-monorepo-toolchain-migration-700-files-lint-typecheck-23x-faster/",
      "source_url": "https://sergioazocar.com/en/blog/how-to-migrate-with-claude-code-and-not-die-trying/",
      "source_name": "Sergio Azócar's blog",
      "source_type": "builder",
      "date": "2026-04-28",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Monorepo size",
          "value": "~1,670 TypeScript files, 7 apps"
        },
        {
          "label": "oxlint PR",
          "value": "700+ files touched"
        },
        {
          "label": "Lint + typecheck",
          "value": "29.04s to 1.25s (23x)"
        },
        {
          "label": "Site redesign",
          "value": "A couple of days (vs ~two weeks estimated)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Migrated the whole site design in a couple of days instead of an estimated two weeks",
          "quote": "Before Claude Code, that same migration would have eaten up two weeks of my time"
        }
      ],
      "techniques": [
        "atomic prs per phase",
        "validate every step",
        "claude.md context",
        "measure before and after",
        "migration guide as input"
      ],
      "caveats": "Personal blog. The speed-ups come from the new tools, not from Claude Code itself; the agent did the code changes. No time or cost figures for the monorepo migration itself.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "nab-cursor-greenfield-payment-app-3-weeks-6000-developers",
      "title": "National Australia Bank builds a greenfield Android payment app in under three weeks with Cursor, against a four-month scope",
      "url": "https://www.shippedbyagents.dev/stories/nab-cursor-greenfield-payment-app-3-weeks-6000-developers/",
      "source_url": "https://cursor.com/blog/nab",
      "source_name": "Cursor customer story",
      "source_type": "vendor",
      "date": "2026-04-23",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Payment app",
          "value": "Under 3 weeks vs. 4-month original scope"
        },
        {
          "label": "Velocity (engineer's claim)",
          "value": "5-8x"
        },
        {
          "label": "Migration pre-development",
          "value": "1 week vs. 2 months planned"
        },
        {
          "label": "BizCalc migration",
          "value": "Expected 2 months vs. 6 months scoped"
        },
        {
          "label": "Assembly migration",
          "value": "3x faster than expected"
        },
        {
          "label": "Developers on Cursor",
          "value": "6,000+, scaling toward 10,000+"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A payment app shipped in three weeks instead of the original four-month scope",
          "metric": "3 weeks vs 4 months",
          "quote": "shipping a hardware-agnostic payment app in three weeks instead of the original four month scope"
        },
        {
          "claim": "The engineer reports a 5 to 8x improvement in development velocity",
          "metric": "5–8x",
          "quote": "We've seen a 5-8x improvement in development velocity."
        },
        {
          "claim": "Legacy modernisations running 3x faster than expected",
          "metric": "3x",
          "quote": "Legacy modernizations are now running 3x faster and at higher quality."
        }
      ],
      "techniques": [
        "cursor plans the work in phases",
        "work split across subagents",
        "model choice per task"
      ],
      "caveats": "Published by Cursor as a customer story. The BizCalc migration was still in progress (an expectation, not a result), and the 3x and 5-8x figures are relative to the bank's own estimates. No code size, cost or quality data.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "spotify-honk-dataset-consumer-migrations-240-prs-10-engineering-weeks",
      "title": "Spotify's Honk agent rolls out 240 dataset migration PRs across downstream pipelines, saving an estimated 10 engineering weeks",
      "url": "https://www.shippedbyagents.dev/stories/spotify-honk-dataset-consumer-migrations-240-prs-10-engineering-weeks/",
      "source_url": "https://engineering.atspotify.com/2026/4/background-coding-agents-dataset-migrations-honk-part-4",
      "source_name": "Spotify Engineering blog (Honk, Part 4)",
      "source_type": "builder",
      "date": "2026-04-22",
      "tools": [
        "Claude"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Automated PRs",
          "value": "240"
        },
        {
          "label": "Scope",
          "value": "~1,800 direct downstream pipelines across two deprecated datasets"
        },
        {
          "label": "Effort saved",
          "value": "An estimated 10 engineering weeks"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "240 automated migration PRs rolled out, saving an estimated 10 engineering weeks",
          "metric": "240 PRs; ~10 engineering weeks saved",
          "quote": "saving an estimated 10 engineering weeks in the process"
        }
      ],
      "techniques": [
        "backstage lineage to find consumers",
        "codesearch",
        "fleetshift orchestration",
        "background coding agent"
      ],
      "caveats": "Company writing about its own platform. Scio pipelines weren't migrated by the agent, and without tests in most repos the agent couldn't verify its changes. The post doesn't say how many of the 240 PRs merged. Publish date read from page metadata.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "intercom-ai-pr-approval-93-percent-agent-driven-prs-19-percent-auto-approved",
      "title": "Intercom lets an AI agent approve 19% of pull requests, with 93% of PRs now agent-driven",
      "url": "https://www.shippedbyagents.dev/stories/intercom-ai-pr-approval-93-percent-agent-driven-prs-19-percent-auto-approved/",
      "source_url": "https://www.intercom.com/blog/ai-is-approving-our-pull-requests-heres-how-we-made-it-safe",
      "source_name": "Intercom blog (Engineering)",
      "source_type": "builder",
      "date": "2026-04-21",
      "tools": [
        "Claude"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Agent-driven PRs",
          "value": "Over 93% (two main codebases)"
        },
        {
          "label": "Auto-approved PRs",
          "value": "Over 19%, no human reviewer"
        },
        {
          "label": "Pilot",
          "value": "100+ PRs, zero reverts"
        },
        {
          "label": "Time-to-approval (p75)",
          "value": "6-16x faster"
        },
        {
          "label": "Fully autonomous PRs",
          "value": "497 in first four weeks of rollout"
        },
        {
          "label": "Revert rate, backend",
          "value": "0.53% AI-authored vs 5.39% human"
        },
        {
          "label": "Revert rate, frontend",
          "value": "0.22% AI-authored vs 2.00% human"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Downtime from breaking code changes dropped 35% while deployments doubled",
          "metric": "-35% downtime",
          "quote": "downtime from breaking code changes dropped 35%, even as our deployments doubled"
        },
        {
          "claim": "6-16x improvement in time-to-approval at the 75th percentile in the pilot",
          "metric": "6-16x"
        },
        {
          "claim": "Zero reverts of AI-approved PRs in the 100+ PR pilot",
          "metric": "0 reverts"
        },
        {
          "claim": "AI-authored backend code reverted at 0.53% vs 5.39% for human-authored",
          "metric": "0.53% vs 5.39%"
        },
        {
          "claim": "Doubled R&D productivity in nine months",
          "quote": "Nine months later, we did it."
        }
      ],
      "techniques": [
        "sub-agent review decomposition",
        "execution path tracing",
        "refuse large prs",
        "engineer feedback flywheel",
        "auditor engagement"
      ],
      "caveats": "Self-reported by Intercom. The post names Claude as writing the code but not the exact tool or the models behind the review agent. Revert-rate comparison is observational: AI-approved PRs are filtered to be small, which may favour them.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cloudflare-opencode-ai-code-review-131k-runs-5169-repos",
      "title": "Cloudflare’s OpenCode-orchestrated AI code review runs 131k times across 5,169 repos in 30 days, median 3m 39s and $0.98",
      "url": "https://www.shippedbyagents.dev/stories/cloudflare-opencode-ai-code-review-131k-runs-5169-repos/",
      "source_url": "https://blog.cloudflare.com/ai-code-review/",
      "source_name": "Cloudflare Blog (Ryan Skidmore)",
      "source_type": "builder",
      "date": "2026-04-20",
      "tools": [
        "OpenCode"
      ],
      "models": [
        "Claude Opus 4.7",
        "GPT-5.4",
        "Claude Sonnet 4.6",
        "GPT-5.3 Codex",
        "Kimi K2.5"
      ],
      "work_type": "New product",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "30-day volume",
          "value": "131,246 review runs on 48,095 MRs across 5,169 repos"
        },
        {
          "label": "Latency",
          "value": "Median 3m 39s per review"
        },
        {
          "label": "Cost",
          "value": "Median $0.98; avg $1.19; P99 $4.45"
        },
        {
          "label": "Findings",
          "value": "159,103 total (~1.2 per review)"
        },
        {
          "label": "Break-glass",
          "value": "288 MRs (0.6%)"
        },
        {
          "label": "Cache hit rate",
          "value": "85.7% on ~120B tokens"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Median review completes in 3 minutes 39 seconds across 131k runs",
          "metric": "median 3m 39s",
          "quote": "the median review completes in 3 minutes and 39 seconds"
        },
        {
          "claim": "Median cost $0.98 per review; break-glass only 0.6% of MRs",
          "metric": "$0.98 median; 0.6% break-glass",
          "quote": "engineers have only needed to “break glass” 288 times (0.6% of merge requests)"
        }
      ],
      "techniques": [
        "specialised reviewer agents",
        "coordinator judge pass",
        "risk tier agent counts",
        "circuit breaker model failback",
        "diff noise filtering"
      ],
      "caveats": "Self-reported Cloudflare engineering blog. This ships an AI review system, not application features; merge-quality impact beyond break-glass rate is not measured. Model names and costs are as of the March–April 2026 window.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "eventship-bubble-rebuild-claude-code-5-weeks-20000-users",
      "title": "Eventship replaces a quoted 6-month, 5-person rebuild with 5 weeks of solo Claude Code work and migrates 20,000 users",
      "url": "https://www.shippedbyagents.dev/stories/eventship-bubble-rebuild-claude-code-5-weeks-20000-users/",
      "source_url": "https://www.gpthacks.com/p/i-replaced-a-6-month-5-person-rebuild",
      "source_name": "GPT Hacks newsletter (Hussein Yahfoufi)",
      "source_type": "builder",
      "date": "2026-04-15",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "5 weeks, solo"
        },
        {
          "label": "Estimate without AI",
          "value": "4-5 engineers for 6 months"
        },
        {
          "label": "Users migrated",
          "value": "About 20,000 users, 500 hosts, 2,000 events"
        },
        {
          "label": "Tool cost",
          "value": "Claude $100/month plan, then $200/month"
        },
        {
          "label": "Parallel sessions",
          "value": "Usually 4"
        }
      ],
      "duration_days": 35,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A solo rebuild in five weeks, against an estimate of 4 to 5 engineers for 6 months",
          "metric": "5 weeks vs 4–5 engineers × 6 months",
          "quote": "the estimate to rebuild the whole thing was around 4-5 engineers and it would take 6 months"
        },
        {
          "claim": "Close to 20,000 users and their data moved over without data loss",
          "metric": "~20,000 users",
          "quote": "close to 20,000 users and all their associated data moved over without data loss"
        }
      ],
      "techniques": [
        "feature docs in notion",
        "about four parallel sessions",
        "manual and automated verification of each feature",
        "rehearsed data migration"
      ],
      "caveats": "Self-reported by the founder. A friend wrote the data migration script. Launch had an outage from unconfigured GCP scaling, plus edge-case bugs afterwards. No line counts or test numbers given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "msr-2026-agent-adoption-open-source-velocity-vs-technical-debt",
      "title": "MSR 2026 study: coding agents give open-source repos a velocity bump only as their first AI tool, but lasting technical debt either way",
      "url": "https://www.shippedbyagents.dev/stories/msr-2026-agent-adoption-open-source-velocity-vs-technical-debt/",
      "source_url": "https://arxiv.org/abs/2601.13597",
      "source_name": "MSR 2026 (Mining Challenge), arXiv 2601.13597",
      "source_type": "independent",
      "date": "2026-04-13",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Commits (agents as first AI tool)",
          "value": "+36.3%"
        },
        {
          "label": "Lines added (agents as first AI tool)",
          "value": "+76.6%"
        },
        {
          "label": "Commits / lines (prior AI IDE users)",
          "value": "+3.1% / −6.3%"
        },
        {
          "label": "Static-analysis warnings",
          "value": "~+18% (persistent)"
        },
        {
          "label": "Cognitive complexity",
          "value": "~+39% (persistent)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Repos where an agent was the first AI tool saw commits up 36.3% and lines added up 76.6%; repos already using AI IDEs saw little change",
          "metric": "+36.3% commits",
          "quote": "AF repos see large average gains"
        }
      ],
      "techniques": [
        "matched difference-in-differences",
        "sonarqube quality metrics",
        "agent-induced technical debt"
      ],
      "caveats": "Observational, open-source repos only, and a Mining Challenge paper on one dataset. Adoption is inferred from the first agent-authored PR, so earlier unlabelled AI use is missed. Replication package: github.com/shyamagarwal13/agentic-coding-impact.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "brightbean-studio-social-media-management-tool-claude-code-codex-three-weeks",
      "title": "A solo developer ships an open-source social media management platform with 12 API integrations in three weeks, with Claude Code building and Codex reviewing",
      "url": "https://www.shippedbyagents.dev/stories/brightbean-studio-social-media-management-tool-claude-code-codex-three-weeks/",
      "source_url": "https://news.ycombinator.com/item?id=47749674",
      "source_name": "Hacker News Show HN (JanSchu)",
      "source_type": "builder",
      "date": "2026-04-13",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [
        "Claude Opus 4.6",
        "Codex 5.3"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Build time",
          "value": "3 weeks (plus 4 full days on specs)"
        },
        {
          "label": "Scope",
          "value": "12 first-party platform API integrations"
        },
        {
          "label": "Parallel agents",
          "value": "Up to 4"
        },
        {
          "label": "Token spend",
          "value": "Roughly even between Claude Code and Codex; frequently hit Claude 5x Max session limits"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A platform he estimates would keep a solo developer busy most of a year, shipped in three weeks",
          "metric": "3 weeks",
          "quote": "The scope would normally keep a solo developer busy for the better part of a year. I shipped it in 3 weeks."
        }
      ],
      "techniques": [
        "specs and architecture doc before code",
        "layers of parallel work streams",
        "git worktree per stream",
        "second model reviews each pr",
        "tests did not catch tenant leaks"
      ],
      "caveats": "Self-reported in a Show HN post and comments. No line counts, user numbers or dollar costs given. He says UX polish took most of the time ('80% of features worked in 20% of the time'). Some commenters questioned AI-written docs and comments.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "zoominfo-angular-to-react-claude-code-15-agents-938-files-one-week",
      "title": "ZoomInfo rewrites a 938-file Angular feature to React in one week with a 15-agent Claude Code swarm, against a six-month 4–5 engineer baseline",
      "url": "https://www.shippedbyagents.dev/stories/zoominfo-angular-to-react-claude-code-15-agents-938-files-one-week/",
      "source_url": "https://engineering.zoominfo.com/experience-with-multi-agent-ai-for-framework-migration-at-zoominfo",
      "source_name": "ZoomInfo Engineering Blog (Faraazuddin Mohammed)",
      "source_type": "builder",
      "date": "2026-04-08",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus",
        "Claude Sonnet"
      ],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Scope",
          "value": "938 files"
        },
        {
          "label": "Agent team",
          "value": "1 Opus lead + 15 Sonnet workers"
        },
        {
          "label": "Agent-driven time",
          "value": "1 week calendar, 8–12 working hours, 1 engineer"
        },
        {
          "label": "Historical baseline",
          "value": "4–5 engineers, 6 months to build the Angular feature"
        }
      ],
      "duration_days": 7,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "938-file Angular-to-React migration done in one week / 8–12 working hours with one engineer and 15 agents",
          "metric": "1 week (8–12 hrs) vs 6 months with 4–5 engineers",
          "quote": "Agent-Driven Teams Mode | 1 week (8-12 working hrs) | 1 | 15"
        }
      ],
      "techniques": [
        "claude code agent teams",
        "shared memory and todo list",
        "agents.md context router",
        "typescript lsp feedback",
        "playwright mcp parity tests"
      ],
      "caveats": "Self-reported by ZoomInfo. The six-month figure is the historical cost to build the Angular feature, not a measured hand port of the same 938 files to React. No dollar cost or production defect counts. The post documents recurring AI slop (test memory leaks, architectural hallucinations, cyclical AST loops) that needed audit agents and human last-mile work.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "atlassian-rovo-dev-code-reviewer-1900-repos-pr-cycle-time-down-30-8-percent",
      "title": "Atlassian's Rovo Dev code review agent cuts median PR cycle time 30.8% across 1,900+ internal repositories in a year-long evaluation",
      "url": "https://www.shippedbyagents.dev/stories/atlassian-rovo-dev-code-reviewer-1900-repos-pr-cycle-time-down-30-8-percent/",
      "source_url": "https://www.atlassian.com/blog/ai-at-work/developer-productivity-improved-with-rovo-dev",
      "source_name": "Inside Atlassian (Andy Wong, Minwoo Jeong)",
      "source_type": "vendor",
      "date": "2026-04-07",
      "tools": [
        "Rovo Dev"
      ],
      "models": [
        "Claude 3.5 Sonnet",
        "gpt-4o-mini",
        "ModernBERT"
      ],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Scope",
          "value": "Over a year across 1,900+ repositories"
        },
        {
          "label": "Median PR cycle time",
          "value": "-30.8%"
        },
        {
          "label": "Human-written review comments",
          "value": "-35.6%"
        },
        {
          "label": "Comment resolution rate",
          "value": "38.70% (human comments 44.45%)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "30.8% reduction in median PR cycle time",
          "metric": "-30.8%",
          "quote": "We observed a significant 30.8% reduction in median PR cycle time."
        },
        {
          "claim": "35.6% fewer human-written review comments",
          "metric": "-35.6%"
        },
        {
          "claim": "38.70% of Rovo Dev comments led directly to code changes",
          "metric": "38.70% code resolution rate"
        }
      ],
      "techniques": [
        "structured zero-shot review prompt",
        "pr and jira context in prompt",
        "llm-as-judge filters hallucinated comments",
        "actionability classifier",
        "human accepts or declines each suggestion"
      ],
      "caveats": "Atlassian sells Rovo Dev, so this is vendor-published, though the evaluation was accepted at ICSE 2026. It is a review agent, not one that writes the code. Cycle-time change is observational over adoption, not a controlled comparison.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "vercel-risk-classifier-auto-merges-58-percent-of-prs-merge-time-29-to-10-9-hours",
      "title": "Vercel auto-merges 58% of PRs in its largest monorepo without human review, cutting average merge time from 29 hours to 10.9",
      "url": "https://www.shippedbyagents.dev/stories/vercel-risk-classifier-auto-merges-58-percent-of-prs-merge-time-29-to-10-9-hours/",
      "source_url": "https://web.archive.org/web/20260407105633/https://vercel.com/blog/58-percent-of-prs-in-our-largest-monorepo-merge-without-human-review",
      "source_name": "Vercel blog (archived copy)",
      "source_type": "vendor",
      "date": "2026-04-06",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs merged without a human reviewer",
          "value": "58%"
        },
        {
          "label": "Average merge time",
          "value": "29 hours → 10.9 hours (-62%)"
        },
        {
          "label": "PR volume",
          "value": "Over 400 PRs a week"
        },
        {
          "label": "Low-risk PRs that skipped review",
          "value": "671, zero reverted"
        },
        {
          "label": "PRs per active author",
          "value": "2.6 → 3.8 a week (+46%)"
        },
        {
          "label": "Time to first review, HIGH-risk large-diff PRs",
          "value": "24.7 hours → 9.0 hours (2.7×)"
        },
        {
          "label": "Deployment rollbacks",
          "value": "2.8 → 1.9 a week"
        },
        {
          "label": "Classifier cost",
          "value": "~$0.054 per assessment (~$51/week)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "An agent merges 58% of PRs without a human reviewer, and average merge time dropped 62% from 29 hours to 10.9 hours",
          "metric": "29 hours → 10.9 hours (-62%)",
          "quote": "Today, an agent reviews and merges 58% of those pull requests without a human reviewer, and average merge time has dropped 62%, from 29 hours to 10.9 hours."
        },
        {
          "claim": "671 low-risk PRs skipped review and none were reverted",
          "metric": "671 PRs, 0 reverts",
          "quote": "671 low-risk PRs skipped review. Zero were reverted."
        },
        {
          "claim": "Individual human throughput up 46%, from 2.6 to 3.8 PRs per author per week",
          "metric": "2.6 → 3.8 PRs per author per week",
          "quote": "Individual human throughput increased 46%. PRs per active author went from 2.6 per week to 3.8 per week."
        },
        {
          "claim": "Time to first review on HIGH-risk, large-diff PRs fell from 24.7 to 9.0 hours",
          "metric": "24.7 hours → 9.0 hours (2.7×)",
          "quote": "Time-to-first-review on HIGH-risk, large-diff PRs dropped from 24.7 hours to 9.0 hours, a 2.7x improvement."
        },
        {
          "claim": "Deployment rollbacks fell from 2.8 to 1.9 a week",
          "metric": "2.8 → 1.9 rollbacks a week",
          "quote": "Deployment rollbacks decreased from 2.8 per week to 1.9 per week."
        }
      ],
      "techniques": [
        "llm risk classifier",
        "auto-approve low-risk prs",
        "prefer false highs over false lows",
        "three-phase rollout with kill switches",
        "misclassified prs added to evals"
      ],
      "caveats": "Vendor-published: Vercel sells Vercel Agent and AI Gateway, which run the workflow. Dead original link: the original vercel.com address (https://vercel.com/blog/58-percent-of-prs-in-our-largest-monorepo-merge-without-human-review) now returns 404 and is not in Vercel's sitemap, so this story links to the Wayback Machine copy of 7 April 2026 (checked 10 November 2026); all numbers were verified from that archived copy, which shows no author names. The source covers an LLM classifier plus auto-approval, not a coding agent writing the code: it does not say which agents (if any) authored the PRs, and it does not name the classifier model (secondary write-ups say Gemini). Self-reported, from an experiment window of unstated length, in a repo dominated by marketing, docs and UI changes, so transfer to riskier codebases is unproven. The control group had 2 reverts out of 513 reviewed low-risk PRs. Cost figures are per-assessment rates, not a total.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "factory-legacy-bench-cobol-java7-fortran",
      "title": "Legacy-Bench: frontier agents pass only 16.9–42.5% of COBOL, Java 7 and Fortran tasks",
      "url": "https://www.shippedbyagents.dev/stories/factory-legacy-bench-cobol-java7-fortran/",
      "source_url": "https://factory.com/news/legacy-bench",
      "source_name": "Factory",
      "source_type": "vendor",
      "date": "2026-04-01",
      "tools": [
        "Factory Droid",
        "OpenAI Codex",
        "Gemini CLI"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Pass rate range (12 model–agent combos)",
          "value": "16.9%–42.5%"
        },
        {
          "label": "Same models on Terminal-Bench 2 / SWE-bench Verified",
          "value": ">70%"
        },
        {
          "label": "Best overall (GPT-5.3-Codex)",
          "value": "42.5%"
        },
        {
          "label": "Tasks no model solves",
          "value": "44 (31 of them COBOL)"
        },
        {
          "label": "Failures where agent believed it succeeded",
          "value": "97%"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "hidden verification tests",
        "legacy language benchmark"
      ],
      "caveats": "Published by Factory, whose own agent (Droid) is reported as outperforming providers' agents. Only 10 sample tasks are public. Models evaluated include GPT-5.3-Codex, Gemini-3.1-Pro, Opus 4.6 and GLM-5.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "jmix-cuba-to-jmix-timesheets-migration-codex-25-hours-vs-215-hour-estimate",
      "title": "Jmix migrates a CUBA Platform Timesheets app to Jmix with Codex in 25 hours against a 215-hour estimate",
      "url": "https://www.shippedbyagents.dev/stories/jmix-cuba-to-jmix-timesheets-migration-codex-25-hours-vs-215-hour-estimate/",
      "source_url": "https://www.jmix.io/blog/migrating-from-cuba-platform-to-jmix-with-ai-agents/",
      "source_name": "Jmix blog (Konstantin Krivopustov)",
      "source_type": "vendor",
      "date": "2026-03-25",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [
        "GPT-5.3-Codex"
      ],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Actual effort",
          "value": "25 man-hours"
        },
        {
          "label": "Migration Advisor estimate",
          "value": "315 man-hours (215 project-specific after a fixed 100-hour baseline)"
        },
        {
          "label": "Source code size",
          "value": "8,062 Java, 2,252 XML, 2,151 SQL lines"
        },
        {
          "label": "Target code size",
          "value": "10,916 Java, 3,513 XML lines"
        },
        {
          "label": "Screens",
          "value": "35 screens, 1 frame"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "The migration took 25 man-hours against a 215-hour project-specific estimate",
          "metric": "25 vs 215 man-hours"
        },
        {
          "claim": "About 10 times less effort than expected",
          "metric": "~10x less effort",
          "quote": "we succeeded with the Timesheets migration and spent about 10 times less effort than expected"
        }
      ],
      "techniques": [
        "plan-first prompts for complex screens",
        "simple screens first",
        "split by package to avoid context overload",
        "agents.md and framework skills",
        "playwright ui testing by the agent"
      ],
      "caveats": "Published by Jmix, which sells the target framework and the migration tooling, about its own sample application, so this is vendor-published. The baseline is Jmix's own Migration Advisor estimate, not a measured manual migration. Some behaviour was changed rather than reproduced (Ctrl+Enter instead of Enter for command input). The author says other projects may not see a similar reduction.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dotnet-runtime-copilot-coding-agent-ten-months-878-prs-535-merged",
      "title": "Ten months of GitHub's Copilot coding agent in dotnet/runtime: 878 PRs, 535 merged, with a revert rate no worse than humans'",
      "url": "https://www.shippedbyagents.dev/stories/dotnet-runtime-copilot-coding-agent-ten-months-878-prs-535-merged/",
      "source_url": "https://devblogs.microsoft.com/dotnet/ten-months-with-cca-in-dotnet-runtime/",
      "source_name": ".NET Blog (Stephen Toub)",
      "source_type": "vendor",
      "date": "2026-03-23",
      "tools": [
        "GitHub Copilot"
      ],
      "models": [
        "Claude Sonnet 4",
        "Claude Sonnet 4.5"
      ],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "CCA PRs in dotnet/runtime",
          "value": "878; 535 merged, 253 closed, 90 open (67.9% success)"
        },
        {
          "label": "Code",
          "value": "Over 95,000 lines added and 31,000 removed"
        },
        {
          "label": "Revert rate",
          "value": "3 of 535 merged (0.6%) vs 0.8% for other PRs"
        },
        {
          "label": "Instructions effect",
          "value": "38.1% success before setup changes, 69% after"
        },
        {
          "label": "Seven repos",
          "value": "2,963 CCA PRs, 1,885 merged (68.6%)"
        },
        {
          "label": "Median time to merge",
          "value": "About 2 days"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 535,
      "claimed_benefits": [
        {
          "claim": "Success rate rose from 41.7% in the first month to about 71%",
          "metric": "41.7% to ~71%"
        },
        {
          "claim": "Good instructions and setup lifted success from 38.1% to 69%",
          "metric": "38.1% to 69%"
        },
        {
          "claim": "Revert rate of merged CCA PRs no higher than other PRs",
          "metric": "0.6% vs 0.8%"
        },
        {
          "claim": "20% of issues tackled were over two years old",
          "metric": "20%"
        }
      ],
      "techniques": [
        "copilot-instructions.md",
        "firewall rules for package feeds",
        "custom skills such as benchmarking bot",
        "assign well-scoped issues",
        "humans push fixes to agent branches"
      ],
      "caveats": "Microsoft owns GitHub, so this is vendor-adjacent, though written by a .NET engineer with detailed data and caveats. CCA tasks were chosen by maintainers, so comparisons with human PRs are between different populations. Compute and CI costs are not analysed. Models named are those used at the time of one experiment.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "eve-wall-e-slack-background-agent-opencode-22-percent-of-merge-requests-53-percent-one-shot",
      "title": "Eve’s Slack-driven background agent WALL-E authors 22% of merge requests in its monorepo, 53% of them one-shot",
      "url": "https://www.shippedbyagents.dev/stories/eve-wall-e-slack-background-agent-opencode-22-percent-of-merge-requests-53-percent-one-shot/",
      "source_url": "https://www.eve.legal/engineering/wall-e-at-eve-an-always-on-agent-that-actually-ships",
      "source_name": "Eve engineering blog (Allen Deng)",
      "source_type": "builder",
      "date": "2026-03-19",
      "tools": [
        "OpenCode"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of merge requests",
          "value": "22% of all merge requests in the monorepo"
        },
        {
          "label": "One-shot rate",
          "value": "53% of WALL-E merge requests need no human iteration"
        },
        {
          "label": "Time per merge request",
          "value": "About 15 minutes"
        },
        {
          "label": "Hackathon to first real merge request",
          "value": "Two weeks"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "WALL-E authors 22% of all merge requests in the monorepo",
          "metric": "22%"
        },
        {
          "claim": "53% of its merge requests are one-shot, with no human iteration",
          "metric": "53%",
          "quote": "53% of those are one-shot"
        },
        {
          "claim": "Merge request ready in about 15 minutes",
          "metric": "about 15 minutes"
        }
      ],
      "techniques": [
        "slack thread per task",
        "full-stack sandbox per thread",
        "browser and database checks before the mr",
        "agent writes lessons to agents.md",
        "maximum agency inside an isolated sandbox"
      ],
      "caveats": "Written by Eve’s own engineer. The post gives no absolute counts of merge requests, team size, cost or time period, and does not name the models used. One-shot means no human iteration before the merge request, as defined by Eve.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "taller-chiron-three-modernization-programs-36-to-9-3-project-weeks-orchestrated-agents",
      "title": "Taller Technologies’ Chiron platform cut three modernization programs from 36 to 9.3 summed project-weeks, with gains coming only once agents were orchestrated",
      "url": "https://www.shippedbyagents.dev/stories/taller-chiron-three-modernization-programs-36-to-9-3-project-weeks-orchestrated-agents/",
      "source_url": "https://arxiv.org/abs/2603.20028",
      "source_name": "arXiv (Maximiliano Armesto and Christophe Kolb, Taller Technologies)",
      "source_type": "builder",
      "date": "2026-03-17",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Summed project duration",
          "value": "36.0 weeks (traditional) to 9.3 weeks (V4), 3.87x faster"
        },
        {
          "label": "Validation-stage issue load",
          "value": "8.03 to 2.09 issues per 100 tasks"
        },
        {
          "label": "First-release coverage",
          "value": "77.0% to 90.5% (task-weighted)"
        },
        {
          "label": "Modeled effort",
          "value": "1,080 to 232.5 person-days"
        },
        {
          "label": "Review-stage containment (V4)",
          "value": "About 51.4% of issues caught before validation"
        },
        {
          "label": "Programs",
          "value": "3: about 30k, 400k and 30k lines"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Summed delivery time cut from 36.0 to 9.3 project-weeks",
          "metric": "3.87x speedup"
        },
        {
          "claim": "Issues reaching validation fell from 8.03 to 2.09 per 100 tasks",
          "metric": "74.0% reduction"
        },
        {
          "claim": "First-release coverage rose from 77.0% to 90.5%",
          "metric": "+13.4 pp"
        }
      ],
      "techniques": [
        "acceptance-criteria validation per task",
        "repository-native pr review",
        "hybrid human-agent execution",
        "agent-generated backlog and documentation",
        "staged delivery: analysis, planning, implementation, validation"
      ],
      "caveats": "Both authors are executives at the company that built Chiron and disclose the conflict of interest. The benchmark was assembled retrospectively from engineering records and practitioner recall, in one organization, with no independent audit. Person-day figures are modeled from staffing assumptions, not logged hours. Models and coding tools are not named.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "datadog-redis-rust-helix-agents-built-full-systems-harness-first",
      "title": "Datadog has Claude Code and Codex build a Redis-compatible server and a Kafka-compatible streaming engine, and leans on test harnesses instead of code review",
      "url": "https://www.shippedbyagents.dev/stories/datadog-redis-rust-helix-agents-built-full-systems-harness-first/",
      "source_url": "https://www.datadoghq.com/blog/ai/harness-first-agents/",
      "source_name": "Datadog blog (Alp Keles, Jai Menon, Sesh Nalla, Vyom Shah)",
      "source_type": "builder",
      "date": "2026-03-09",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [
        "Claude Opus 4.5"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "redis-rust memory",
          "value": "87% reduction after agent-guided optimisation (started at 8x Redis 8.4's memory)"
        },
        {
          "label": "redis-rust latency",
          "value": "Comparable to Redis in production-like staging"
        },
        {
          "label": "Helix simulation",
          "value": "500 seeds per component, then 10 million seeds across all components"
        },
        {
          "label": "Helix throughput",
          "value": "About 93% of peak disk throughput"
        },
        {
          "label": "Spec overhead",
          "value": "Writing invariants takes roughly 2-3x the effort of writing the code"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "redis-rust reached production-like staging with comparable latency and an 87% memory reduction after agent-guided iteration",
          "metric": "-87% memory",
          "quote": "redis-rust reached production-like staging with comparable latency and an 87% memory reduction after agent-guided iteration"
        },
        {
          "claim": "Producing to Helix averaged 22.2 ms against 116 ms for the baseline Kafka cluster",
          "metric": "22.2 ms vs 116 ms",
          "quote": "Producing to Helix averaged 22.2 ms, compared to 116 ms for the baseline Kafka cluster."
        }
      ],
      "techniques": [
        "harness-first engineering",
        "deterministic simulation testing",
        "checks replace line-by-line review",
        "agent reads production metrics"
      ],
      "caveats": "Company writing about its own internal projects; redis-rust reached staging, not full production. First post in a series, so later results may differ. Datadog sells observability, which the approach leans on.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "composio-agent-orchestrator-built-by-its-own-agents-8-days-80-prs",
      "title": "A developer's agent orchestrator built most of itself in eight days, merging 80+ of its own PRs from up to 16 parallel Claude Code sessions",
      "url": "https://www.shippedbyagents.dev/stories/composio-agent-orchestrator-built-by-its-own-agents-8-days-80-prs/",
      "source_url": "https://news.ycombinator.com/item?id=47219229",
      "source_name": "Hacker News Show HN (prateekk77)",
      "source_type": "builder",
      "date": "2026-03-02",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Build time",
          "value": "8 days"
        },
        {
          "label": "Commits",
          "value": "134"
        },
        {
          "label": "Self-merged PRs",
          "value": "80+"
        },
        {
          "label": "Peak parallel agents",
          "value": "About 16 on the same repo"
        },
        {
          "label": "GitHub stars",
          "value": "2,700+ in two weeks"
        }
      ],
      "duration_days": 8,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Got to 13 PRs in a day with the first hacky version",
          "metric": "13 PRs in a day"
        },
        {
          "claim": "Merged 80+ of its own PRs in 8 days",
          "metric": "80+ PRs, 8 days"
        }
      ],
      "techniques": [
        "git worktree per agent",
        "orchestrator agent reads backlog",
        "merge queue rebases downstream agents",
        "ci logs fed back to agent",
        "agent addresses review comments"
      ],
      "caveats": "Short self-reported Show HN post from a company that sells agent tooling; the fuller write-up is on X. No cost, line counts or quality data. The tool also supports Codex and Aider, but the post only says Claude Code sessions were used. Stack options (tmux or Docker, GitHub or Linear) are pluggable, so none is listed.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "metr-2026-uplift-update-selection-effects",
      "title": "METR's late-2025 follow-up: raw data hints at speed-up, but developers refusing to work without AI breaks the study",
      "url": "https://www.shippedbyagents.dev/stories/metr-2026-uplift-update-selection-effects/",
      "source_url": "https://metr.org/blog/2026-02-24-uplift-update/",
      "source_name": "METR",
      "source_type": "independent",
      "date": "2026-02-24",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Developers",
          "value": "57"
        },
        {
          "label": "Repositories",
          "value": "143"
        },
        {
          "label": "Tasks",
          "value": "800+"
        },
        {
          "label": "Original-cohort estimate",
          "value": "−18% time (CI −38% to +9%)"
        },
        {
          "label": "New-cohort estimate",
          "value": "−4% time (CI −15% to +9%)"
        },
        {
          "label": "Developers withholding tasks from no-AI condition",
          "value": "30–50% (survey)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "randomised controlled trial",
        "selection effects",
        "multiple agents running concurrently"
      ],
      "caveats": "METR itself calls this very weak evidence because of selection effects; pay was also reduced from $150/hr to $50/hr.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cloudflare-vinext-one-engineer-rebuilds-nextjs-on-vite-in-a-week-1100-dollars-tokens",
      "title": "Cloudflare engineer rebuilds the Next.js API surface on Vite in under a week with Claude in OpenCode, for about $1,100 in tokens",
      "url": "https://www.shippedbyagents.dev/stories/cloudflare-vinext-one-engineer-rebuilds-nextjs-on-vite-in-a-week-1100-dollars-tokens/",
      "source_url": "https://blog.cloudflare.com/vinext/",
      "source_name": "Cloudflare Blog (Steve Faulkner)",
      "source_type": "builder",
      "date": "2026-02-24",
      "tools": [
        "OpenCode"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Under a week; first commit 13 February"
        },
        {
          "label": "Token cost",
          "value": "About $1,100 in Claude API tokens"
        },
        {
          "label": "Agent sessions",
          "value": "Over 800 sessions in OpenCode"
        },
        {
          "label": "People",
          "value": "One engineer (technically engineering manager) directing AI"
        },
        {
          "label": "Tests",
          "value": "1,700+ Vitest tests and 380 Playwright E2E tests"
        },
        {
          "label": "API coverage",
          "value": "94% of the Next.js 16 API surface"
        },
        {
          "label": "Production build time",
          "value": "Up to 4.4x faster than Next.js 16 (Vite 8 / Rolldown, 33-route test app)"
        },
        {
          "label": "Client bundle size",
          "value": "Up to 57% smaller gzipped"
        }
      ],
      "duration_days": null,
      "cost_usd": 1100,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Rebuilt the Next.js API surface in under a week, against months or years for a team",
          "metric": "Under a week",
          "quote": "This time we did it in under a week. One engineer (technically engineering manager) directing AI."
        },
        {
          "claim": "Whole project cost about $1,100 in tokens",
          "metric": "About $1,100",
          "quote": "The whole thing cost about $1,100 in tokens."
        },
        {
          "claim": "Production builds up to 4x faster than Next.js in early benchmarks",
          "metric": "Up to 4.4x faster (Vite 8 / Rolldown)"
        },
        {
          "claim": "Client bundles up to 57% smaller",
          "metric": "57% smaller gzipped"
        },
        {
          "claim": "94% of the Next.js 16 API surface covered",
          "metric": "94%",
          "quote": "Coverage sits at 94% of the Next.js 16 API surface."
        }
      ],
      "techniques": [
        "plan the architecture first",
        "tests ported from the next.js suite",
        "merge when tests pass",
        "agent code review",
        "agent-browser for browser-level testing"
      ],
      "caveats": "Written by the engineer who did the work, on the company blog, at launch. vinext is labelled experimental and \"not yet been battle-tested with any meaningful traffic at scale\". Benchmarks use a single 33-route app and measure compilation and bundling, not serving performance; the post says to treat them as directional. Static pre-rendering at build time was not supported at launch. The post says some AI PRs \"were just wrong\" and needed regular course correction. Exact model versions are not named.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "ladybird-libjs-cpp-to-rust-port-claude-code-codex-two-weeks",
      "title": "Ladybird ports its JavaScript engine's parser and bytecode compiler from C++ to Rust in two weeks with Claude Code and Codex",
      "url": "https://www.shippedbyagents.dev/stories/ladybird-libjs-cpp-to-rust-port-claude-code-codex-two-weeks/",
      "source_url": "https://ladybird.org/posts/adopting-rust/",
      "source_name": "Ladybird blog (Andreas Kling)",
      "source_type": "builder",
      "date": "2026-02-23",
      "tools": [
        "Claude Code",
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Output size",
          "value": "About 25,000 lines of Rust"
        },
        {
          "label": "Time",
          "value": "About two weeks"
        },
        {
          "label": "Estimate without AI",
          "value": "Multiple months by hand (author's estimate)"
        },
        {
          "label": "test262",
          "value": "52,898 tests, 0 regressions"
        },
        {
          "label": "Ladybird regression tests",
          "value": "12,461 tests, 0 regressions"
        }
      ],
      "duration_days": 14,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 25,000 lines of Rust in about two weeks, where Kling says it would have taken him multiple months by hand",
          "metric": "~2 weeks vs multiple months",
          "quote": "The same work would have taken me multiple months to do by hand."
        },
        {
          "claim": "Zero regressions across 52,898 test262 tests and 12,461 Ladybird regression tests, and no performance regressions",
          "metric": "0 regressions",
          "quote": "No performance regressions on any of the JS benchmarks we track either."
        }
      ],
      "techniques": [
        "human-directed, not autonomous",
        "hundreds of small prompts",
        "adversarial review by different models",
        "byte-for-byte identical output",
        "lockstep mode running both pipelines"
      ],
      "caveats": "Self-reported by the project founder. Not autonomous: the author decided what to port, in what order and how the Rust should look. The code is intentionally non-idiomatic Rust that mirrors the C++. No cost or token figures given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "soundsignal-solo-developer-permit-intelligence-platform-five-days-claude-code-95-prs-506-tests",
      "title": "W. David Betts builds SoundSignal, a full-stack building permit intelligence platform, in five days with Claude Code: 95 PRs, about 29,500 lines and 506 tests",
      "url": "https://www.shippedbyagents.dev/stories/soundsignal-solo-developer-permit-intelligence-platform-five-days-claude-code-95-prs-506-tests/",
      "source_url": "https://wdbetts.com/writing/building-soundsignal-5-days-claude-code/",
      "source_name": "wdbetts.com (W. David Betts)",
      "source_type": "builder",
      "date": "2026-02-22",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "5 calendar days"
        },
        {
          "label": "Claude Code sessions",
          "value": "56 (414 messages)"
        },
        {
          "label": "Commits and PRs",
          "value": "143 commits, 95 pull requests"
        },
        {
          "label": "Code",
          "value": "About 29,500 lines"
        },
        {
          "label": "Tests",
          "value": "506 (464 backend, 42 frontend)"
        },
        {
          "label": "Infrastructure code",
          "value": "4,400+ lines of Terraform for 15+ AWS services"
        },
        {
          "label": "Estimated Claude Code cost",
          "value": "About $200"
        }
      ],
      "duration_days": 5,
      "cost_usd": 200,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Full-stack platform with AWS deployment built in five days",
          "metric": "5 days"
        },
        {
          "claim": "Straightforward features often went from request to deployed in under 10 minutes",
          "metric": "under 10 minutes",
          "quote": "Getting from “I want X” to “X is deployed” often took under 10 minutes for straightforward features."
        },
        {
          "claim": "Claude Code wrote tests without being asked, reaching 506 tests",
          "metric": "506 tests"
        }
      ],
      "techniques": [
        "git worktree per task",
        "every change through a pr",
        "claude.md as project memory",
        "one feature per session",
        "time-box exploration in claude.md"
      ],
      "caveats": "Self-reported by the builder. The developer worked full-time and says he knew what he wanted to build. He ran Claude Code with --dangerously-skip-permissions throughout and says he would not on a production codebase with secrets. The post does not report users or revenue. The cost is his estimate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stripe-minions-unattended-coding-agents-1300-merged-prs-a-week-no-human-written-code",
      "title": "Stripe's unattended Minions agents produce over 1,300 merged pull requests a week, human-reviewed but with no human-written code",
      "url": "https://www.shippedbyagents.dev/stories/stripe-minions-unattended-coding-agents-1300-merged-prs-a-week-no-human-written-code/",
      "source_url": "https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2",
      "source_name": "Stripe Dev Blog (Alistair Gray)",
      "source_type": "builder",
      "date": "2026-02-19",
      "tools": [
        "Goose"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Merged PRs",
          "value": "Over 1,300 a week, completely minion-produced (up from 1,000 at Part 1)"
        },
        {
          "label": "Environment",
          "value": "Prewarmed devboxes usable in about 10 seconds"
        },
        {
          "label": "Test suite",
          "value": "Over three million tests"
        },
        {
          "label": "CI attempts",
          "value": "One full CI run plus at most one repair attempt, then hand to a human"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Over 1,300 merged pull requests a week are completely minion-produced, up from 1,000",
          "metric": "1,300+ merged PRs/week",
          "quote": "Over 1,300 Stripe pull requests (up from 1,000 as of Part 1) merged each week are completely minion-produced, human-reviewed, but containing no human-written code."
        }
      ],
      "techniques": [
        "prewarmed disposable devboxes",
        "blueprints mix deterministic steps and agent loops",
        "shift feedback left with local lint",
        "capped ci retries then human handoff",
        "full permissions inside quarantined sandbox"
      ],
      "caveats": "Self-reported. No merge rate, revert rate or time-saved figures. Weekly PR counts lack a denominator of total Stripe PRs. Goose is a Stripe-internal fork; Cursor and Claude Code are used separately by engineers and are not part of Minions.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "openai-harness-engineering-codex-million-lines-zero-manual-code",
      "title": "OpenAI team ships an internal product with zero hand-written code: ~1M lines and ~1,500 merged PRs in five months with Codex",
      "url": "https://www.shippedbyagents.dev/stories/openai-harness-engineering-codex-million-lines-zero-manual-code/",
      "source_url": "https://openai.com/index/harness-engineering/",
      "source_name": "OpenAI engineering blog",
      "source_type": "vendor",
      "date": "2026-02-11",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Duration",
          "value": "~5 months"
        },
        {
          "label": "Codebase size",
          "value": "On the order of 1 million lines (incl. infra, tooling, docs)"
        },
        {
          "label": "PRs merged",
          "value": "~1,500"
        },
        {
          "label": "Team",
          "value": "3 engineers, growing to 7"
        },
        {
          "label": "Throughput",
          "value": "~3.5 PRs per engineer per day"
        },
        {
          "label": "Hand-written code",
          "value": "0 lines"
        },
        {
          "label": "Time vs hand-coding",
          "value": "~1/10th (team's estimate)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": 1500,
      "claimed_benefits": [
        {
          "claim": "Built in about a tenth of the time it would have taken to write by hand (OpenAI's estimate)",
          "metric": "~1/10th the time",
          "quote": "We estimate that we built this in about 1/10th the time it would have taken to write the code by hand."
        },
        {
          "claim": "About a million lines and roughly 1,500 merged pull requests in five months, averaging 3.5 PRs per engineer per day",
          "metric": "3.5 PRs/engineer/day",
          "quote": "This translates to an average throughput of 3.5 PRs per engineer per day"
        }
      ],
      "techniques": [
        "no manually-written code",
        "agents.md as a table of contents",
        "custom linters with remediation instructions",
        "agent-to-agent review",
        "recurring cleanup agents"
      ],
      "caveats": "Written by the vendor about its own tool. Line count includes docs and tooling. No cost, time-vs-estimate or quality figures.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "anthropic-16-agents-build-c-compiler-20000-dollars",
      "title": "16 parallel Claude agents build a 100,000-line Rust C compiler that builds Linux, for ~$20,000",
      "url": "https://www.shippedbyagents.dev/stories/anthropic-16-agents-build-c-compiler-20000-dollars/",
      "source_url": "https://www.anthropic.com/engineering/building-c-compiler",
      "source_name": "Anthropic Engineering blog",
      "source_type": "vendor",
      "date": "2026-02-05",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 4.6"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "API cost",
          "value": "Just under $20,000"
        },
        {
          "label": "Duration",
          "value": "~2 weeks, nearly 2,000 Claude Code sessions"
        },
        {
          "label": "Output",
          "value": "100,000-line compiler"
        },
        {
          "label": "Tokens",
          "value": "2B input, 140M output"
        },
        {
          "label": "Test-suite pass rate",
          "value": "99% on most suites incl. GCC torture tests"
        }
      ],
      "duration_days": 14,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A 100,000-line C compiler that builds Linux 6.9 on x86, ARM and RISC-V, for just under $20,000 in API costs over about two weeks",
          "metric": "~$20,000; ~2 weeks",
          "quote": "Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V."
        },
        {
          "claim": "99% pass rate on most compiler test suites, including the GCC torture tests",
          "metric": "99%",
          "quote": "has a 99% pass rate on most compiler test suites including the GCC torture test suite"
        }
      ],
      "techniques": [
        "16 parallel agents",
        "each agent in its own container",
        "coordination through git and lock files",
        "no orchestrator",
        "test harness with a fast random sample"
      ],
      "caveats": "Vendor-published capability stress test, not a product. Calls out to GCC for 16-bit x86 real-mode boot. Its own assembler and linker are still buggy (the demo used GCC's). It isn't a drop-in compiler, and its output is less efficient than GCC with optimisations off. Rust quality is 'nowhere near' expert level. Late on, new features often broke existing ones. The author says he's uneasy about deploying software nobody has personally verified.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "instacart-caper-jetpack-compose-migration-ai-5-7x-faster-300-hours-saved",
      "title": "Instacart's Caper team migrates its smart-cart Android app to Jetpack Compose with AI, 5-7x faster on navigation",
      "url": "https://www.shippedbyagents.dev/stories/instacart-caper-jetpack-compose-migration-ai-5-7x-faster-300-hours-saved/",
      "source_url": "https://company.instacart.com/tech-innovation/migrating-to-jetpack-compose-how-ai-accelerated-our-journey-at-caper",
      "source_name": "Instacart tech blog (Matt Kranzler)",
      "source_type": "builder",
      "date": "2026-02-03",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Phase 2 speed-up",
          "value": "5-7x"
        },
        {
          "label": "Engineering hours saved",
          "value": "Estimated 300-350 (Phase 2)"
        },
        {
          "label": "Navigation scope",
          "value": "30+ sub-navigation graphs, 130+ destinations"
        },
        {
          "label": "Fragment features to migrate",
          "value": "100+"
        },
        {
          "label": "Migration guide",
          "value": "325+ lines, 17-step skill"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "5-7x speed increase in Phase 2, saving an estimated 300-350 engineering hours",
          "metric": "5-7x; 300-350 hours",
          "quote": "The 5–7x speed increase we observed in Phase 2 saved an estimated 300–350 engineering hours"
        },
        {
          "claim": "AI reduced variance across migrations, improving quality and predictability"
        },
        {
          "claim": "Migrations ran in parallel with feature development without blocking it"
        }
      ],
      "techniques": [
        "git history as context",
        "migration guide as code",
        "ai skill",
        "screenshot parity checks",
        "phased migration"
      ],
      "caveats": "No coding tool or model is named. Speed-up and hours saved are estimates for Phase 2 only; Phase 3 was still in progress and Phase 4 had just started when published.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "fullscript-nitro-background-agent-19-percent-of-prs-in-three-months-ceo-ships-feature",
      "title": "Fullscript's Nitro background agent creates 19% of pull requests three months after launch on a decade-old Rails monolith",
      "url": "https://www.shippedbyagents.dev/stories/fullscript-nitro-background-agent-19-percent-of-prs-in-three-months-ceo-ships-feature/",
      "source_url": "https://builders.fullscript.com/posts/lessons-learned-from-building-nitro-fullscripts-autonomous-background-agent",
      "source_name": "Fullscript Engineering (Nick Crum)",
      "source_type": "builder",
      "date": "2026-02-03",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of PRs created by Nitro",
          "value": "19% in January 2026, three months after launch"
        },
        {
          "label": "Review usage",
          "value": "Tagged as reviewer on 45% of PRs, voluntarily"
        },
        {
          "label": "Build time",
          "value": "About half a year from idea to usable"
        },
        {
          "label": "Codebase",
          "value": "Decade-old Rails monolith, millions of lines, 180+ engineers"
        },
        {
          "label": "CEO contributions",
          "value": "Three changes, including a mobile app feature"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "19% of pull requests created by Nitro three months after launch",
          "metric": "19% of PRs",
          "quote": "In January, three months later, 19% of our pull requests were created by this tool"
        },
        {
          "claim": "Developers voluntarily tag Nitro to review 45% of pull requests",
          "metric": "45% of PRs",
          "quote": "developers are voluntarily tagging the agent to code review on 45% of our pull requests"
        },
        {
          "claim": "Retroactive reviews of past PRs that caused production issues found the bugs"
        }
      ],
      "techniques": [
        "fresh environment per request",
        "trigger from issue tracker and code review",
        "agent addresses accumulated pr feedback",
        "document conventions as intent layer",
        "human review on every pr"
      ],
      "caveats": "Self-reported. No absolute PR counts, merge rates or quality metrics beyond one reverted N+1 fix that caused production 500s. Every PR still needs human review, CI and E2E tests.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dropbox-cursor-550k-file-monorepo-1m-lines-accepted-monthly",
      "title": "Dropbox indexes a 550,000-file monorepo in Cursor and accepts more than a million agent-generated lines a month",
      "url": "https://www.shippedbyagents.dev/stories/dropbox-cursor-550k-file-monorepo-1m-lines-accepted-monthly/",
      "source_url": "https://cursor.com/blog/dropbox",
      "source_name": "Cursor blog (Dropbox customer story)",
      "source_type": "vendor",
      "date": "2026-01-26",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Monorepo indexed",
          "value": "550,000+ files"
        },
        {
          "label": "Accepted Cursor lines",
          "value": "1M+ per month"
        },
        {
          "label": "Engineering adoption",
          "value": "90%+ of engineers use AI tools weekly"
        },
        {
          "label": "Traffic context",
          "value": "300,000+ requests per second from Dropbox data centers"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than one million Cursor-generated lines accepted every month",
          "metric": "1M+ lines/month",
          "quote": "Dropbox's engineers now accept more than one million lines of AI-suggested code every month with Cursor."
        },
        {
          "claim": "More than 90% of engineers use AI tools weekly",
          "metric": "90%+",
          "quote": "More than 90 percent of Dropbox engineers now use AI tools weekly, with Cursor as a primary driver of that activity."
        }
      ],
      "techniques": [
        "monorepo semantic indexing",
        "organic champion-led adoption",
        "single-click signup friction removal",
        "leadership dogfooding at hackathon"
      ],
      "caveats": "Vendor-published Cursor customer story. Exact PR-throughput and cycle-time numbers are not given (only “upper tier of industry benchmarks”). Lines “accepted” are not the same as merged production code.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "saru-solo-developer-200k-line-multi-tenant-saas-three-months-claude-code",
      "title": "Solo developer builds a 200,000-line multi-tenant SaaS with four portals and four APIs in three months with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/saru-solo-developer-200k-line-multi-tenant-saas-three-months-claude-code/",
      "source_url": "https://ko-chan.github.io/saru-blog/posts-en/saru-claude-code-200k-lines-6/",
      "source_name": "Saru Blog (ko-chan)",
      "source_type": "builder",
      "date": "2026-01-24",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Lines of code",
          "value": "~200K (Go 84K + TypeScript 113K)"
        },
        {
          "label": "Time",
          "value": "~3 months (Oct 2025–)"
        },
        {
          "label": "Commits",
          "value": "311"
        },
        {
          "label": "Developers",
          "value": "1"
        },
        {
          "label": "Time per feature",
          "value": "2-3 days → 4-6 hours"
        },
        {
          "label": "Daily commits",
          "value": "2-3 → 5-10"
        },
        {
          "label": "Test coverage",
          "value": "Always 80%+"
        },
        {
          "label": "Human time",
          "value": "2-3 hours per day"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Time per feature cut from 2-3 days to 4-6 hours",
          "metric": "2-3 days → 4-6 hours"
        },
        {
          "claim": "Feels 3-5x faster overall",
          "metric": "3-5x",
          "quote": "Feels 3-5x faster."
        },
        {
          "claim": "Test coverage always 80%+",
          "metric": "Always 80%+"
        },
        {
          "claim": "Adding an entity across 4 APIs and 4 portals takes 30 minutes instead of half a day",
          "metric": "Half a day → 30 minutes"
        }
      ],
      "techniques": [
        "spec.md then plan.md before implementation",
        "human reviews every change",
        "human security review",
        "reject features not needed now",
        "claude writes tests after implementation"
      ],
      "caveats": "Self-reported by an anonymous solo developer (ko-chan). Before/after comparisons are against the author's pre-AI experience, not a measured baseline, and the overall speed-up is described as how it \"feels\". No users, revenue or production deployment are reported. Cost is given only as tens to hundreds of dollars a month.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "msr-2026-why-33000-agent-pull-requests-get-rejected",
      "title": "MSR 2026 study of 33,596 agent-written GitHub pull requests: 71% merged, and the rejects are mostly ignored, duplicated or failing CI",
      "url": "https://www.shippedbyagents.dev/stories/msr-2026-why-33000-agent-pull-requests-get-rejected/",
      "source_url": "https://arxiv.org/abs/2601.15195",
      "source_name": "arXiv 2601.15195 (MSR 2026)",
      "source_type": "independent",
      "date": "2026-01-21",
      "tools": [
        "OpenAI Codex",
        "GitHub Copilot",
        "Devin",
        "Cursor",
        "Claude Code"
      ],
      "models": [],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs studied",
          "value": "33,596 across five agents"
        },
        {
          "label": "Overall merge rate",
          "value": "71.48%"
        },
        {
          "label": "Merge rate by agent",
          "value": "Codex 82.59%, Cursor 65.22%, Claude Code 59.04%, Devin 53.76%, Copilot 43.04%"
        },
        {
          "label": "Top rejection reason",
          "value": "Abandoned or never reviewed (228 of 600 sampled)"
        },
        {
          "label": "Effect of CI failures",
          "value": "Each failed CI check cuts merge odds by about 15%"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "rejection taxonomy",
        "aidev-pop dataset",
        "logistic regression on pr outcomes"
      ],
      "caveats": "Short MSR Mining Challenge paper. Sample sizes per agent are very uneven (Codex 21,799 PRs vs. Claude Code 459). Measures merge outcomes, not code quality or productivity. Studies the same AIDev dataset as the existing MSR 2026 adoption story.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cisco-codex-build-optimization-defect-remediation",
      "title": "Cisco uses Codex across 15+ repos to cut build times ~20% and speed up C/C++ defect fixing",
      "url": "https://www.shippedbyagents.dev/stories/cisco-codex-build-optimization-defect-remediation/",
      "source_url": "https://openai.com/index/cisco/",
      "source_name": "OpenAI customer story",
      "source_type": "vendor",
      "date": "2026-01-20",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Engineering hours saved",
          "value": "1,500+ per month"
        },
        {
          "label": "Build time",
          "value": "~20% reduction"
        },
        {
          "label": "Defect resolution throughput",
          "value": "10–15× increase"
        },
        {
          "label": "Repositories analyzed",
          "value": "15+"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 20% off build times and more than 1,500 engineering hours saved a month across 15+ repositories",
          "metric": "~20% faster builds; 1,500+ hours/month",
          "quote": "a ~20% reduction in build times and more than 1,500 engineering hours saved per month"
        },
        {
          "claim": "Defect fixing that took weeks of manual effort now takes hours, a 10 to 15x increase in defect resolution throughput",
          "metric": "10–15x",
          "quote": "What once took weeks of manual effort now completes in hours"
        },
        {
          "claim": "Splunk's React 18 to 19 UI migrations went from weeks to days",
          "quote": "compressing weeks of work into days"
        },
        {
          "claim": "More than 95% of new AI features written by Codex",
          "metric": "95%+"
        }
      ],
      "techniques": [
        "autonomous compile-test-fix loops",
        "codex generates and follows a plan document",
        "build log and dependency graph analysis"
      ],
      "caveats": "Vendor-published.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-01-20 from the date shown on the source page (read via Wayback, as the live page returned a Cloudflare challenge)."
    },
    {
      "slug": "jason-hoffman-judoka-app-claude-code-19-days-309000-lines",
      "title": "Judoka, a dojo app for iOS, Android and web, goes from start to App Store approval in 19 days with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/jason-hoffman-judoka-app-claude-code-19-days-309000-lines/",
      "source_url": "https://fullhoffman.com/2026/01/16/on-building-software-with-claude-code/",
      "source_name": "fullhoffman.com (Jason Hoffman)",
      "source_type": "builder",
      "date": "2026-01-16",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "19 active development days to App Store approval"
        },
        {
          "label": "Size",
          "value": "About 309,000 lines of code"
        },
        {
          "label": "Tests",
          "value": "2,726"
        },
        {
          "label": "Database migrations",
          "value": "214"
        },
        {
          "label": "Commits",
          "value": "1,488 across three phases (533 + 652 + 303)"
        },
        {
          "label": "Share written by Claude",
          "value": "About 70% of the implementation"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A production app in 19 active development days; Hoffman says size estimates put it at 9,500 to 12,000 person-hours, or 12 to 15 months for a five-person team",
          "metric": "19 days vs 9,500–12,000 person-hours",
          "quote": "That corresponds to five or six years of solo work, or 12–15 months for a five-person team."
        }
      ],
      "techniques": [
        "human designs architecture and specifies features",
        "claude writes most of the implementation"
      ],
      "caveats": "Self-reported. The 9,500-12,000 person-hour comparison is a COCOMO-style estimate, and the author calls the ~71x figure a throughput comparison, not a productivity study. The post itself is labelled as AI-generated writing directed by Jason Hoffman. No cost figures; maintenance cost unmeasured.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cursor-fastrender-browser-hundreds-of-agents",
      "title": "Cursor's agent swarm writes a 1M+ line browser in about a week, and critics find it barely compiles",
      "url": "https://www.shippedbyagents.dev/stories/cursor-fastrender-browser-hundreds-of-agents/",
      "source_url": "https://cursor.com/blog/scaling-agents",
      "source_name": "Cursor blog",
      "source_type": "vendor",
      "date": "2026-01-14",
      "tools": [
        "Cursor"
      ],
      "models": [
        "GPT-5.2",
        "GPT-5.1-Codex"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Duration",
          "value": "Close to a week"
        },
        {
          "label": "Output",
          "value": "1M+ lines across ~1,000 files"
        },
        {
          "label": "Agents",
          "value": "Hundreds of concurrent workers"
        },
        {
          "label": "Solid → React migration (same post)",
          "value": "3+ weeks, +266K/-193K edits"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Agents wrote over 1 million lines across 1,000 files in close to a week",
          "metric": "1M+ lines in ~1 week",
          "quote": "The agents ran for close to a week, writing over 1 million lines of code across 1,000 files."
        },
        {
          "claim": "A long-running agent made video rendering 25x faster",
          "metric": "25x",
          "quote": "A long-running agent made video rendering 25x faster with an efficient Rust version."
        }
      ],
      "techniques": [
        "hundreds of concurrent agents",
        "planners, workers and a judge",
        "long-running agents"
      ],
      "caveats": "Vendor research experiment, and the browser was a test objective for the harness, not a product. The Register (26 Jan 2026) reported failing GitHub Actions on main, commits that didn't build cleanly, pages taking about a minute to load, and reliance on existing projects. A Servo maintainer called the code 'a tangle of spaghetti'. The JS engine was vendored from the engineer's own personal parser project. No cost was disclosed. Lin says the harness deliberately tolerated a steady rate of compile errors to keep throughput up. Cursor says Opus 4.5 stopped early and took shortcuts compared with GPT-5.2.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "ramp-inspect-background-agent-30-percent-merged-prs-no-mandate",
      "title": "Ramp's in-house background agent Inspect writes about 30% of merged PRs in its frontend and backend repos, without a mandate",
      "url": "https://www.shippedbyagents.dev/stories/ramp-inspect-background-agent-30-percent-merged-prs-no-mandate/",
      "source_url": "https://engineering.ramp.com/post/why-we-built-our-background-agent",
      "source_name": "Ramp Engineering blog",
      "source_type": "builder",
      "date": "2026-01-12",
      "tools": [
        "OpenCode"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Share of merged PRs",
          "value": "About 30% of PRs merged to frontend and backend repos written by Inspect"
        },
        {
          "label": "Time to that level",
          "value": "A couple of months"
        },
        {
          "label": "Repo image freshness",
          "value": "Rebuilt every 30 minutes"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 30% of all pull requests merged to Ramp's frontend and backend repos are written by Inspect, a couple of months in, with no mandate",
          "metric": "~30% of merged PRs",
          "quote": "We didn't force anyone to use Inspect"
        }
      ],
      "techniques": [
        "sandboxed vm per session",
        "full dev environment in the sandbox",
        "wired into sentry and datadog",
        "repo image rebuilt every 30 minutes",
        "background agent"
      ],
      "caveats": "Company writing about its own tool. No figures for PR size, review effort, defects or cost. Says it supports all frontier models without naming which ones are used most.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2026-01-12 from the page's datePublished meta tag."
    },
    {
      "slug": "simon-willison-ports-justhtml-to-javascript-codex-4-hours",
      "title": "Simon Willison ports an HTML5 parser from Python to JavaScript with Codex CLI in about 4.5 hours",
      "url": "https://www.shippedbyagents.dev/stories/simon-willison-ports-justhtml-to-javascript-codex-4-hours/",
      "source_url": "https://simonwillison.net/2025/Dec/15/porting-justhtml/",
      "source_name": "simonwillison.net",
      "source_type": "builder",
      "date": "2025-12-15",
      "tools": [
        "OpenAI Codex"
      ],
      "models": [
        "GPT-5.2"
      ],
      "work_type": "Language port",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Idea to finished library",
          "value": "~4–4.5 hours"
        },
        {
          "label": "Prompts",
          "value": "8"
        },
        {
          "label": "Output",
          "value": "9,000 lines of JavaScript, 43 commits"
        },
        {
          "label": "html5lib tests passing",
          "value": "9,200 (tokenizer 6,810/6,810; tree 1,770/1,782)"
        },
        {
          "label": "Estimated API-price cost",
          "value": "$29.41 (covered by $20/month ChatGPT Plus)"
        },
        {
          "label": "Tokens",
          "value": "1.46M input + 97.1M cached input, 625,563 output"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "9,000 lines of fully tested JavaScript across 43 commits, about 4 hours from idea to finished library",
          "metric": "~4 hours",
          "quote": "ended up producing 9,000 lines of fully tested JavaScript across 43 commits"
        },
        {
          "claim": "Tokens worth an estimated $29.41 at API prices, covered by a $20/month ChatGPT Plus plan",
          "metric": "$29.41"
        }
      ],
      "techniques": [
        "agent writes a spec from the existing code first",
        "small first milestone",
        "existing conformance test suite",
        "commit and push often",
        "minimal supervision"
      ],
      "caveats": "Straight port of an existing library against a comprehensive existing test suite: close to ideal conditions. Ran with --yolo (no sandbox or approvals). Hit the Plus plan's five-hour token window and paused ~40 minutes. The author raises open questions on copyright, ethics and whether code built this way should be trusted in production.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "emil-stenstrom-justhtml-python-html5-parser-copilot-agent",
      "title": "Emil Stenström builds JustHTML, a pure-Python HTML5 parser passing all 9,200+ html5lib tests, with Copilot agent mode",
      "url": "https://www.shippedbyagents.dev/stories/emil-stenstrom-justhtml-python-html5-parser-copilot-agent/",
      "source_url": "https://simonwillison.net/2025/Dec/14/justhtml/",
      "source_name": "Simon Willison (simonwillison.net), summarising Emil Stenström's write-up",
      "source_type": "builder",
      "date": "2025-12-14",
      "tools": [
        "GitHub Copilot"
      ],
      "models": [
        "Claude Sonnet 3.7",
        "Gemini 3 Pro",
        "Claude Opus"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "A couple of months of off-hours"
        },
        {
          "label": "Implementation code",
          "value": "~3,000 lines (plus ~11,000 lines of tests)"
        },
        {
          "label": "html5lib-tests passing",
          "value": "All 9,200+"
        },
        {
          "label": "Test coverage",
          "value": "100%"
        },
        {
          "label": "Commits",
          "value": "1,215 (per Simon Willison's follow-up post)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A pure-Python HTML5 parser of about 3,000 lines that passes all 9,200+ html5lib tests, with 100% test coverage",
          "metric": "9,200+ tests",
          "quote": "Passes all 9,200+ tests in the official html5lib-tests suite"
        }
      ],
      "techniques": [
        "conformance test suite from the start",
        "human chose the api",
        "agent did the typing",
        "vibe engineering, not vibe coding"
      ],
      "caveats": "Fact source is Simon Willison's write-up quoting Emil's own blog post (Emil's original wasn't reachable when researched; worth linking directly before publishing). Heavy human engineering throughout: Emil chose the design, restarted from scratch once, and says 'the agent did the typing; I did the thinking'. Problem was picked because a comprehensive test suite already existed. How much was translation of html5ever versus original work is unclear.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "vellum-15-engineers-double-monthly-prs-946-to-1910-with-background-and-review-agents",
      "title": "Vellum's 15 engineers double monthly PRs from 946 to 1,910 with background agents and Codex as first-pass reviewer",
      "url": "https://www.shippedbyagents.dev/stories/vellum-15-engineers-double-monthly-prs-946-to-1910-with-background-and-review-agents/",
      "source_url": "https://www.vellum.ai/blog/how-we-use-coding-agents-to-2x-engineering-output",
      "source_name": "Vellum Blog (Nicolas Zeeb)",
      "source_type": "builder",
      "date": "2025-12-12",
      "tools": [
        "Devin",
        "OpenAI Codex",
        "Cursor",
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "PRs",
          "value": "946 in May → 1,910 in October"
        },
        {
          "label": "Team",
          "value": "15 engineers, no added headcount"
        },
        {
          "label": "Background-agent PRs",
          "value": "37% of PRs now initiated by background agents"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "PRs doubled from 946 to 1,910 a month with the same team",
          "metric": "946 (May) → 1,910 (October)",
          "quote": "we've enabled our engineers to 2x the amount of PRs we shipped from May (946) to October (1910)."
        },
        {
          "claim": "37% of PRs initiated by background agents",
          "metric": "37%",
          "quote": "This resulted in 37% of our PRs now being initiated by background agents"
        }
      ],
      "techniques": [
        "background agents from linear and slack",
        "codex as first-pass code review",
        "post-merge human review",
        "one-line make commands",
        "preview environments for every pr"
      ],
      "caveats": "Self-reported on Vellum's company blog, which also promotes a webinar with Cognition (maker of Devin). PR counts compare one month with another; PR size and quality are not measured. The year of the May baseline is 2024 per the post; the October figure's year is not stated explicitly. Humans review post-merge rather than before merge, so Codex review is the main pre-merge gate.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "lee-robinson-cursor-com-cms-to-code-migration-three-days-260-dollars",
      "title": "Lee Robinson migrates cursor.com off a headless CMS back to code and Markdown in three days with Cursor agents for $260",
      "url": "https://www.shippedbyagents.dev/stories/lee-robinson-cursor-com-cms-to-code-migration-three-days-260-dollars/",
      "source_url": "https://leerob.com/agents",
      "source_name": "leerob.com (Lee Robinson)",
      "source_type": "builder",
      "date": "2025-12",
      "tools": [
        "Cursor"
      ],
      "models": [
        "Claude Opus 4.5"
      ],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Calendar time",
          "value": "3 days (estimated 1–2 weeks)"
        },
        {
          "label": "Token cost",
          "value": "$260.32 (297.4M tokens, mostly cached)"
        },
        {
          "label": "Agent requests",
          "value": "344 agent requests, 66 Tab changes, 67 commits"
        },
        {
          "label": "Diff",
          "value": "+43K / −322K lines"
        },
        {
          "label": "Prior CMS CDN spend",
          "value": "$56,848 since September launch"
        }
      ],
      "duration_days": 3,
      "cost_usd": 260.32,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "CMS-to-code migration finished in three days against a 1–2 week estimate",
          "metric": "3 days vs 1–2 weeks",
          "quote": "I had estimated it would take 1-2 weeks, but was able to finish the migration in three days with $260 in tokens and hundreds of agents."
        },
        {
          "claim": "Build times 2x faster after cutting CMS network I/O at prerender",
          "metric": "2x faster builds",
          "quote": "build times are 2x faster by cutting out network I/O going to the CMS when prerendering pages."
        }
      ],
      "techniques": [
        "api export scripts from cms",
        "browser screenshot parity loop",
        "subagents for parallel refactors",
        "delete abstraction complexity",
        "content as code"
      ],
      "caveats": "Written by a Cursor employee about migrating Cursor’s own marketing site with Cursor. CDN savings and “2x faster builds” are claimed after the migration but without a controlled before/after study. December 2025 publish month only (day not stated).",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "spotify-honk-background-agent-1500-merged-prs",
      "title": "Spotify's background coding agent merges 1,500+ migration PRs with 60–90% time savings",
      "url": "https://www.shippedbyagents.dev/stories/spotify-honk-background-agent-1500-merged-prs/",
      "source_url": "https://engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1",
      "source_name": "Spotify Engineering",
      "source_type": "builder",
      "date": "2025-11-06",
      "tools": [
        "Claude Code",
        "Claude Agent SDK"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Agent PRs merged to production",
          "value": "1,500+"
        },
        {
          "label": "Time saved on these migrations",
          "value": "60–90% vs writing by hand"
        },
        {
          "label": "Users",
          "value": "Hundreds of developers"
        },
        {
          "label": "Pre-agent baseline",
          "value": "Maven updater script grew to 20,000+ lines"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 1,500 agent pull requests merged into production",
          "metric": "1,500+ PRs",
          "quote": "our agents have generated more than 1,500 pull requests that teams across Spotify have merged into our production codebase"
        },
        {
          "claim": "60 to 90% time saved on these migrations compared with writing the code by hand",
          "metric": "60–90%",
          "quote": "a total time saving of 60–90% compared to writing the code by hand"
        }
      ],
      "techniques": [
        "background coding agent",
        "natural-language prompts instead of migration scripts",
        "prs across thousands of repos",
        "fleet management"
      ],
      "caveats": "Self-reported, and the time-saving method isn't detailed. Spotify flags slow and unpredictable agent runs, the need for new validation, safety guardrails and significant LLM cost. Later posts say prompts still evolve 'mostly by intuition'. A June 2026 Spotify post reports 76% more PRs to review, so review has become the bottleneck. Anthropic also publishes a co-marketing case study that claims 'up to 90%'.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2025-11-06 from the visible page date \"Nov 6, 2025\"."
    },
    {
      "slug": "mitchell-hashimoto-ghostty-update-feature-amp-16-dollars",
      "title": "Mitchell Hashimoto ships a non-trivial Ghostty feature with Amp in 16 sessions for $15.98",
      "url": "https://www.shippedbyagents.dev/stories/mitchell-hashimoto-ghostty-update-feature-amp-16-dollars/",
      "source_url": "https://mitchellh.com/writing/non-trivial-vibing",
      "source_name": "mitchellh.com",
      "source_type": "builder",
      "date": "2025-10-11",
      "tools": [
        "Amp"
      ],
      "models": [],
      "work_type": "Feature",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Agent sessions",
          "value": "16"
        },
        {
          "label": "Token spend",
          "value": "$15.98"
        },
        {
          "label": "Hands-on time (author's estimate)",
          "value": "~8 hours, over 3 days from first to last commit"
        }
      ],
      "duration_days": 3,
      "cost_usd": 15.98,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A shipped feature for $15.98 in tokens across 16 sessions and about 8 hours of the author's time",
          "metric": "$15.98; ~8 hours",
          "quote": "The work took a total of 16 separate sessions totalling $15.98 in token spend on Amp"
        }
      ],
      "techniques": [
        "plan first",
        "agent prototypes the ui",
        "anti-slop cleanup sessions",
        "published all agent sessions",
        "final manual review"
      ],
      "caveats": "A lot of human coding too: he edited after almost every session and did a final manual review before merging. The agent couldn't fix the key titlebar bug (he pivoted to an overlay), and its first backend was 'obviously the wrong approach' and thrown away. He thinks it was faster than doing it alone, but no estimate was made in advance. The author says he has no financial ties to any AI company.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dmitry-brant-modernizes-25-year-old-ftape-kernel-driver-claude-code",
      "title": "Claude Code ports a 25-year-old Linux tape driver from kernel 2.4 to 6.8 over two evenings",
      "url": "https://www.shippedbyagents.dev/stories/dmitry-brant-modernizes-25-year-old-ftape-kernel-driver-claude-code/",
      "source_url": "https://dmitrybrant.com/2025/09/07/using-claude-code-to-modernize-a-25-year-old-kernel-driver",
      "source_name": "dmitrybrant.com",
      "source_type": "builder",
      "date": "2025-09-07",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Time",
          "value": "Two evenings"
        },
        {
          "label": "Kernel version jump",
          "value": "2.4 → 6.8"
        },
        {
          "label": "Author's by-hand estimate",
          "value": "Probably several weeks (would have meant learning 25-year-old kernel development)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "The abandoned ftape driver running on a modern kernel over two evenings, where the author thinks doing it himself would probably have taken several weeks",
          "metric": "2 evenings vs several weeks",
          "quote": "was suddenly completed over the course of two evenings"
        },
        {
          "claim": "The author says he learned modern kernel conventions along the way",
          "quote": "I did end up learning a huge amount of things that will be actually useful to me in the future"
        }
      ],
      "techniques": [
        "iterating on compiler output",
        "out-of-tree build",
        "pasting dmesg logs for diagnosis"
      ],
      "caveats": "Needed several manual code fixes and back-and-forth. Hardware testing needed sudo, so the human loaded modules and pasted logs by hand. The author stresses it wouldn't have been possible without baseline kernel-module and C knowledge. Hobby project, single developer; the by-hand estimate is his own guess.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "uber-ureview-ai-code-review-90-percent-of-65000-weekly-diffs-75-percent-useful",
      "title": "Uber's uReview AI code reviewer covers over 90% of ~65,000 weekly diffs, with 75% of comments rated useful",
      "url": "https://www.shippedbyagents.dev/stories/uber-ureview-ai-code-review-90-percent-of-65000-weekly-diffs-75-percent-useful/",
      "source_url": "https://www.uber.com/us/en/blog/ureview/",
      "source_name": "Uber Engineering blog (Sonal Mahajan, Shauvik Roy Choudhary, Will Bond)",
      "source_type": "builder",
      "date": "2025-08-12",
      "tools": [
        "Unstated"
      ],
      "models": [
        "Claude 4 Sonnet",
        "OpenAI o4-mini-high"
      ],
      "work_type": "Company rollout",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Coverage",
          "value": "Over 90% of ~65,000 weekly diffs landed"
        },
        {
          "label": "Comments rated useful",
          "value": "75% (by engineers who interact with the tool)"
        },
        {
          "label": "Comments addressed",
          "value": "Over 65% (vs 51% for human-written comments)"
        },
        {
          "label": "Review latency",
          "value": "Median 4 minutes per commit in CI"
        },
        {
          "label": "Commits processed",
          "value": "Over 10,000 a week, excluding config files"
        },
        {
          "label": "Estimated time saved",
          "value": "~1,500 hours a week (~39 developer years a year)"
        },
        {
          "label": "Monorepos",
          "value": "All six"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Analyses over 90% of the ~65,000 diffs landed each week",
          "metric": "Over 90% of ~65,000 weekly diffs",
          "quote": "uReview today analyzes over 90% of the weekly ~65,000 diffs (equivalent of pull requests) landed at Uber."
        },
        {
          "claim": "75% of comments marked useful and over 65% addressed",
          "metric": "75% useful, 65%+ addressed",
          "quote": "Engineers who interact with the tool mark 75% of its comments as useful, and we see over 65% of its posted comments addressed."
        },
        {
          "claim": "Addressed at a higher rate than human reviewers' comments (51%)",
          "metric": "65% vs 51%",
          "quote": "Internal audits show that only 51% of human-written comments are considered as bugs by the author and addressed in the same changeset."
        },
        {
          "claim": "An estimated 1,500 hours saved a week",
          "metric": "~1,500 hours a week (estimate)",
          "quote": "This translates to approximately 1,500 hours saved weekly, equivalent to nearly 39 developer years annually."
        }
      ],
      "techniques": [
        "prompt chaining",
        "specialized assistants",
        "confidence-score grading",
        "suppress low-value comment categories",
        "gradual rollout"
      ],
      "caveats": "Self-reported by Uber. This is an AI code reviewer, not an agent that writes code, and the post gives no merge-rate, cycle-time or defect outcomes. It does not say which coding agents authored the reviewed diffs, so the tool is Unstated. Published August 2025, earlier than Uber's other stories on the site. Volume figures are inconsistent: '~65,000' diffs is called weekly in one place and 'per month' in another, and 'over 90%' coverage sits awkwardly with 'over 10,000 commits' processed a week. The 1,500 hours a week is an estimate assuming a second human reviewer needs 10 minutes per commit. The usefulness rate counts only engineers who rate comments. Fixer, the companion tool that proposes code changes, is not covered.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "puzzmo-orta-therox-six-weeks-claude-code-tech-debt-and-missing-link",
      "title": "Puzzmo’s Orta Therox clears years of tech-debt migrations in six weeks with Claude Code and ships Missing Link from a prototype monorepo",
      "url": "https://www.shippedbyagents.dev/stories/puzzmo-orta-therox-six-weeks-claude-code-tech-debt-and-missing-link/",
      "source_url": "https://blog.puzzmo.com/posts/2025/07/30/six-weeks-of-claude-code/",
      "source_name": "Puzzmo Blog (Orta Therox)",
      "source_type": "builder",
      "date": "2025-07-30",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Window",
          "value": "6 weeks of Claude Code use"
        },
        {
          "label": "Conversations",
          "value": "~147 conversations across 19 repos/projects"
        },
        {
          "label": "Prototype-to-admin game",
          "value": "Couple of hours for a designer; Missing Link released that way"
        },
        {
          "label": "Side-project blog tooling",
          "value": "Adium-theme Claude chat viewer estimated 10–12 hours by hand, ~2 hours thinking time with Claude"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Years of tech-debt / tech-innovation backlog done in just over a month and a half as side work on the existing roadmap",
          "quote": "This was years of \"tech debt\" / \"tech innovation\" backlog for me! Done in just over a month and a half."
        },
        {
          "claim": "Game designer can go from idea to something running on puzzmo.com for admins in a couple of hours",
          "metric": "couple of hours",
          "quote": "a game designer can go from an idea to something running on puzzmo.com for admins in a couple of hours"
        }
      ],
      "techniques": [
        "two clones two vscode profiles",
        "prototype monorepo for games",
        "github action draft prs in triage",
        "write tests then delete to explore",
        "claude yolo permissions"
      ],
      "caveats": "Self-reported retrospective. Author notes PR/commit/LOC charts did not jump despite felt pace increase, and cites research that developers overestimate AI impact. No per-migration timings, costs, or defect counts. Missing Link is described as an experiment that created product-process tension about keeping prototype code in production.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "metr-2025-rct-experienced-oss-developers-slower",
      "title": "METR RCT: experienced open-source developers took 19% longer with early-2025 AI tools",
      "url": "https://www.shippedbyagents.dev/stories/metr-2025-rct-experienced-oss-developers-slower/",
      "source_url": "https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/",
      "source_name": "METR",
      "source_type": "independent",
      "date": "2025-07-10",
      "tools": [
        "Cursor"
      ],
      "models": [
        "Claude 3.5/3.7 Sonnet"
      ],
      "work_type": "Research / benchmark",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Developers",
          "value": "16"
        },
        {
          "label": "Tasks",
          "value": "246"
        },
        {
          "label": "Measured effect",
          "value": "19% longer with AI"
        },
        {
          "label": "Expected speed-up (before)",
          "value": "24%"
        },
        {
          "label": "Believed speed-up (after)",
          "value": "20%"
        },
        {
          "label": "Repo size (average)",
          "value": "22K+ stars, 1M+ lines"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [],
      "techniques": [
        "randomised controlled trial",
        "issues randomly assigned to ai-allowed or ai-disallowed"
      ],
      "caveats": "METR now marks these results as out of date; see its February 2026 update.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "indragie-karunaratne-context-macos-app-claude-code",
      "title": "Solo developer ships a 20,000-line native macOS app, writing fewer than 1,000 lines by hand",
      "url": "https://www.shippedbyagents.dev/stories/indragie-karunaratne-context-macos-app-claude-code/",
      "source_url": "https://www.indragie.com/blog/i-shipped-a-macos-app-built-entirely-by-claude-code",
      "source_name": "indragie.com",
      "source_type": "builder",
      "date": "2025-07-01",
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Sonnet 4",
        "Claude Opus 4"
      ],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Codebase size",
          "value": "~20,000 lines"
        },
        {
          "label": "Written by hand",
          "value": "<1,000 lines (author's estimate)"
        },
        {
          "label": "Release automation script",
          "value": "2,000 lines of Python, written by Claude in a few hours of iteration"
        },
        {
          "label": "Tool cost",
          "value": "$200/month"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Of 20,000 lines, the author estimates he wrote fewer than 1,000 by hand",
          "metric": "<1,000 of 20,000 lines",
          "quote": "of the 20,000 lines of code in this project, I estimate that I wrote less than 1,000 lines by hand"
        },
        {
          "claim": "His first shipped side project in six years",
          "metric": "first in 6 years"
        },
        {
          "claim": "A release script that saves tens of minutes of manual work on every release",
          "quote": "This script will save me tens of minutes of manual work on every release I publish"
        },
        {
          "claim": "Feels like an extra 5 hours a day for $200 a month",
          "metric": "$200/month",
          "quote": "It's like I found an extra 5 hours every day, and all it cost me was $200 a month."
        }
      ],
      "techniques": [
        "detailed feature specs",
        "automated release script",
        "claude debugs its own script"
      ],
      "caveats": "No total build time given. Claude struggled with Swift Concurrency, tended to reach for legacy AppKit/Obj-C APIs, and couldn't drive the app UI to debug, so logs and screenshots had to be pasted in by hand. Needed detailed specs and planning to get past prototype quality. The author says he has no affiliation with Anthropic.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "cloudflare-workers-oauth-provider-claude-code-five-days",
      "title": "Cloudflare's Kenton Varda prompts an OAuth 2.1 provider library out of Claude Code in about five days",
      "url": "https://www.shippedbyagents.dev/stories/cloudflare-workers-oauth-provider-claude-code-five-days/",
      "source_url": "https://news.ycombinator.com/item?id=44174405",
      "source_name": "Hacker News comments by Kenton Varda (Cloudflare)",
      "source_type": "builder",
      "date": "2025-06-03",
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time with AI",
          "value": "~5 days of heavy work (with other duties), plus minor commits over the month"
        },
        {
          "label": "Library size",
          "value": "~5,000 lines incl. comments, tests and docs"
        },
        {
          "label": "Author's by-hand estimate",
          "value": "2–5x longer (\"a few weeks, maybe months\")"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A ~5,000-line library prompted out in about five days; the author estimates it would have taken 2x to 5x longer by hand",
          "metric": "2–5x",
          "quote": "So I estimate it would have taken me 2x-5x longer to write this library by hand."
        }
      ],
      "techniques": [
        "prompts kept in commit history",
        "every line reviewed",
        "cross-checked against the rfcs"
      ],
      "caveats": "The author calls it a near-ideal case: a well-known standard on a well-known platform with a clear spec. He says AI hasn't saved him much time on the Workers runtime itself. Two vulnerabilities (CVE-2025-4143, missing redirect_uri check; CVE-2025-4144, PKCE bypass) were fixed in v0.0.5 in May 2025. The advisories don't say whether the AI or a human introduced them. The time saving is the author's own guess.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "pieter-levels-flight-sim-cursor-1m-arr-17-days",
      "title": "Pieter Levels builds a browser flight sim with Cursor in ~3 hours; it hits $1M ARR in 17 days",
      "url": "https://www.shippedbyagents.dev/stories/pieter-levels-flight-sim-cursor-1m-arr-17-days/",
      "source_url": "https://x.com/levelsio/status/1899596115210891751",
      "source_name": "X (@levelsio)",
      "source_type": "builder",
      "date": "2025-03-11",
      "tools": [
        "Cursor"
      ],
      "models": [],
      "work_type": "New product",
      "codebase": "greenfield",
      "key_numbers": [
        {
          "label": "Time to first playable version",
          "value": "~3 hours (author's estimate)"
        },
        {
          "label": "Revenue after 17 days",
          "value": "$87,000 MRR (~$1M ARR)"
        },
        {
          "label": "Players after 17 days",
          "value": "320,000"
        },
        {
          "label": "Size when Cursor started struggling",
          "value": "3,000 lines in one file"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "$87,000 MRR ($1M ARR) and 320,000 players 17 days after launch",
          "metric": "$87,000 MRR",
          "quote": "Revenue update: $87,000 MRR (which is $1M ARR)"
        },
        {
          "claim": "The author says without AI it would have taken him 10 to 100x more time",
          "metric": "10–100x",
          "quote": "Without AI it would have taken me 10-100x more time"
        }
      ],
      "techniques": [
        "single-prompt start",
        "building in public",
        "asking the ai to check for security holes"
      ],
      "caveats": "Self-reported on X. Revenue is from in-game ads and sponsors and owes a lot to the author's large existing audience, so it says little about the code. By his own account the build 'didn't go 100% smooth', with about 80% going well. A single large file slowed edits (15 seconds per change) and caused Cursor issues. MRR at a launch-hype peak isn't sustained revenue. Third-party posts also credit Grok 3 and other tools.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "airbnb-llm-enzyme-to-rtl-3500-files",
      "title": "Airbnb migrates ~3.5K Enzyme test files to React Testing Library in 6 weeks with LLMs",
      "url": "https://www.shippedbyagents.dev/stories/airbnb-llm-enzyme-to-rtl-3500-files/",
      "source_url": "https://airbnb.tech/infrastructure/accelerating-large-scale-test-migration-with-llms/",
      "source_name": "Airbnb Engineering & Data Science blog",
      "source_type": "builder",
      "date": "2025-03",
      "tools": [
        "Unstated"
      ],
      "models": [],
      "work_type": "Test migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Test files migrated",
          "value": "~3,500"
        },
        {
          "label": "Actual duration",
          "value": "6 weeks"
        },
        {
          "label": "Manual estimate",
          "value": "1.5 years of engineering time"
        },
        {
          "label": "Auto-migrated in first bulk run",
          "value": "75% in 4 hours"
        },
        {
          "label": "After 4 days of prompt tuning",
          "value": "97%"
        },
        {
          "label": "Prompt size",
          "value": "40,000–100,000 tokens"
        }
      ],
      "duration_days": 42,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 3,500 test files migrated in 6 weeks, against an estimate of 1.5 years of engineering time by hand",
          "metric": "6 weeks vs 1.5 years",
          "quote": "We’d originally estimated this would take 1.5 years of engineering time to do by hand, but — using a combination of frontier models and robust automation — we finished the entire migration in just 6 weeks."
        },
        {
          "claim": "75% of target files migrated automatically in the first four-hour bulk run",
          "metric": "75% in 4 hours",
          "quote": "we successfully migrated 75% of our target files in just four hours"
        },
        {
          "claim": "97% migrated after four days of prompt and script tuning",
          "metric": "97%",
          "quote": "After four days of prompt and script refinement using the “sample, tune, and sweep” strategy, we reached 97% of the 3.5K original Enzyme files."
        }
      ],
      "techniques": [
        "per-file state machine",
        "validation steps",
        "retry loops",
        "dynamic prompting with validation errors",
        "sample, tune, and sweep"
      ],
      "caveats": "Model names and dollar cost not disclosed. Published date is approximate (March 2025, per InfoQ coverage).",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "google-ads-int32-to-int64-id-migration",
      "title": "Google Ads uses LLMs to migrate 32-bit IDs to 64-bit in a 500M+ line codebase",
      "url": "https://www.shippedbyagents.dev/stories/google-ads-int32-to-int64-id-migration/",
      "source_url": "https://arxiv.org/abs/2501.06972",
      "source_name": "Google experience report (arXiv 2501.06972)",
      "source_type": "builder",
      "date": "2025-01-12",
      "tools": [],
      "models": [],
      "work_type": "Refactor / maintenance",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase size",
          "value": "500M+ lines"
        },
        {
          "label": "Code changes fully AI-authored (landed CLs)",
          "value": "80%"
        },
        {
          "label": "Total migration time",
          "value": "~50% reduction (engineer-reported estimate)"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "80% of the code changes in landed changelists were fully AI-authored",
          "metric": "80%",
          "quote": "We found that 80% of the code modifications in the landed CLs were fully AI-authored"
        },
        {
          "claim": "Total migration time cut by an estimated 50%, as reported by the engineers doing it",
          "metric": "~50%",
          "quote": "The total time spent on the migration was reduced by an estimated 50% as reported by the engineers doing the migration"
        }
      ],
      "techniques": [
        "code search and kythe to find locations",
        "fine-tuned gemini on internal code",
        "changes must pass unit tests",
        "changes sharded to code owners for review"
      ],
      "caveats": "Google uses its own model; authors call it an experience report, not a controlled study.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2025-01-12 from the arXiv submission date on abs/2501.06972."
    },
    {
      "slug": "google-junit3-to-junit4-migration",
      "title": "Google migrates 5,359 test files from JUnit3 to JUnit4 in 3 months with LLMs",
      "url": "https://www.shippedbyagents.dev/stories/google-junit3-to-junit4-migration/",
      "source_url": "https://arxiv.org/abs/2501.06972",
      "source_name": "Google experience report (arXiv 2501.06972)",
      "source_type": "builder",
      "date": "2025-01-12",
      "tools": [],
      "models": [],
      "work_type": "Test migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Files migrated",
          "value": "5,359"
        },
        {
          "label": "Lines modified",
          "value": "149,000+"
        },
        {
          "label": "Duration",
          "value": "3 months"
        },
        {
          "label": "AI-generated code committed unchanged",
          "value": "~87%"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "5,359 files migrated, changing more than 149,000 lines, in 3 months",
          "metric": "5,359 files in 3 months",
          "quote": "With this technique we were able to migrate 5,359 files modifying more than 149,000 lines of code in 3 months."
        },
        {
          "claim": "About 87% of the AI-generated code was committed without any change",
          "metric": "~87%"
        }
      ],
      "techniques": [
        "prompt as a list of human migration rules",
        "fine-tuned gemini on internal code",
        "small changelists sent to test owners",
        "limited weekly change volume to avoid overwhelming reviewers"
      ],
      "caveats": "Experience report; Google uses its own model.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2025-01-12 from the arXiv submission date on abs/2501.06972."
    },
    {
      "slug": "google-joda-time-to-java-time-migration",
      "title": "Google estimates ~89% time saved on Joda-Time → java.time migration clusters",
      "url": "https://www.shippedbyagents.dev/stories/google-joda-time-to-java-time-migration/",
      "source_url": "https://arxiv.org/abs/2501.06972",
      "source_name": "Google experience report (arXiv 2501.06972)",
      "source_type": "builder",
      "date": "2025-01-12",
      "tools": [],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Time saved on small clusters (estimate)",
          "value": "~89%"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "About 89% of the time humans would have needed saved on small clusters (Google's current estimate)",
          "metric": "~89%"
        }
      ],
      "techniques": [
        "clustering changes with kythe",
        "related files migrated in one prompt",
        "build, test and fix failures automatically"
      ],
      "caveats": "Ongoing migration; estimate comes from tech leads comparing against the manual approach, extrapolated from a sample of changelists.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2025-01-12 from the arXiv submission date on abs/2501.06972."
    },
    {
      "slug": "nubank-devin-etl-monolith-migration-12x",
      "title": "Nubank uses fine-tuned Devin to split a 6M-line ETL monolith, claiming 8–12x engineering-hour efficiency",
      "url": "https://www.shippedbyagents.dev/stories/nubank-devin-etl-monolith-migration-12x/",
      "source_url": "https://devin.ai/customers/nubank",
      "source_name": "Cognition (devin.ai customer story)",
      "source_type": "vendor",
      "date": "2024-11-14",
      "tools": [
        "Devin"
      ],
      "models": [],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Codebase",
          "value": "6M+ lines, ~100,000 data-class implementations"
        },
        {
          "label": "Original estimate",
          "value": "18 months, 1,000+ engineers"
        },
        {
          "label": "Engineering-hour efficiency",
          "value": "8–12x"
        },
        {
          "label": "Cost savings on delegated scope",
          "value": "20x+"
        },
        {
          "label": "Time per sub-task after fine-tuning",
          "value": "~40 → ~10 minutes"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "12x efficiency in engineering hours saved and over 20x cost savings on the migrations handed to Devin",
          "metric": "12x hours; 20x+ cost",
          "quote": "achieve a 12x efficiency improvement in terms of engineering hours saved, and over 20x cost savings"
        },
        {
          "claim": "Time per sub-task dropped from about 40 minutes to 10 after fine-tuning",
          "metric": "40 → 10 min",
          "quote": "Roughly 40 minutes per sub-task dropped to 10"
        },
        {
          "claim": "Task completion scores doubled and task speed improved 4x after fine-tuning",
          "metric": "2x; 4x",
          "quote": "we observed a doubling of Devin’s task completion scores after fine-tuning, as well as a 4x improvement in task speed"
        }
      ],
      "techniques": [
        "fine-tuned on earlier manual migrations",
        "many instances in parallel",
        "engineers review and merge prs"
      ],
      "caveats": "Vendor-published. The page states 8x, 8–12x and 12x in different places. Efficiency compares typical manual hours with hours spent prompting and reviewing. Cost compares Devin's run cost with engineer hourly cost against an estimated manual baseline. The work was highly repetitive and suited to agents. Needed custom fine-tuning, and early weeks had frequent errors to fix.",
      "updated": "2026-11-16",
      "update_note": "Set published to 2024-11-14 from the page's JSON-LD datePublished (replacing the December 2024 LinkedIn announcement month)."
    },
    {
      "slug": "amazon-q-java-17-upgrades-4500-developer-years",
      "title": "Amazon upgrades tens of thousands of production apps to Java 17 with Amazon Q Developer",
      "url": "https://www.shippedbyagents.dev/stories/amazon-q-java-17-upgrades-4500-developer-years/",
      "source_url": "https://aws.amazon.com/blogs/devops/amazon-q-developer-just-reached-a-260-million-dollar-milestone",
      "source_name": "AWS DevOps & Developer Productivity Blog",
      "source_type": "vendor",
      "date": "2024-08-01",
      "tools": [
        "Amazon Q Developer"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Applications upgraded",
          "value": "Tens of thousands (Java 8/11 → 17)"
        },
        {
          "label": "Development work saved (estimate)",
          "value": "4,500+ years"
        },
        {
          "label": "Annual cost savings (estimate)",
          "value": "$260M from performance improvements"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Over 4,500 years of development work saved compared with manual upgrades",
          "metric": "4,500+ developer-years",
          "quote": "a savings of over 4,500 years of development work"
        },
        {
          "claim": "Performance improvements worth $260 million a year, estimated from hosts removed after the Java 17 upgrade",
          "metric": "$260M a year",
          "quote": "performance improvements worth $260 million dollars in annual cost savings"
        }
      ],
      "techniques": [
        "code transformation agent",
        "fleet-wide java version upgrade"
      ],
      "caveats": "Vendor (Amazon) reporting on its own product; both figures are Amazon's estimates, described as conservative.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stripe-claude-code-rollout-1370-engineers",
      "title": "Stripe rolls Claude Code out to 1,370 engineers with a zero-configuration binary",
      "url": "https://www.shippedbyagents.dev/stories/stripe-claude-code-rollout-1370-engineers/",
      "source_url": "https://claude.com/customers/stripe",
      "source_name": "Anthropic (claude.com customer story)",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Engineers with Claude Code enabled",
          "value": "1,370"
        },
        {
          "label": "Time to produce signed enterprise binary",
          "value": "2–3 months"
        },
        {
          "label": "Coding assistants enabled at Stripe",
          "value": "6"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Claude Code set up for 1,370 engineers with nothing to configure",
          "metric": "1,370 engineers",
          "quote": "Deployed Claude Code to 1,370 engineers with a zero-configuration setup that works immediately on every developer’s machine"
        },
        {
          "claim": "One team migrated 10,000 lines of Scala to Java in four days, against an estimate of ten engineering weeks by hand",
          "metric": "4 days vs 10 engineering weeks",
          "quote": "a team used Claude models to migrate 10,000 lines of Scala to Java in four days, a project estimated at ten engineering weeks by hand"
        }
      ],
      "techniques": [
        "zero-configuration rollout",
        "signed enterprise binary",
        "pre-installed on every laptop and dev box"
      ],
      "caveats": "Vendor-published. Stripe states it hasn't isolated metrics attributable to any single AI assistant.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "stripe-scala-to-java-migration-4-days",
      "title": "Stripe migrates 10,000 lines of Scala to Java in four days",
      "url": "https://www.shippedbyagents.dev/stories/stripe-scala-to-java-migration-4-days/",
      "source_url": "https://claude.com/customers/stripe",
      "source_name": "Anthropic (claude.com customer story)",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Claude"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Lines migrated",
          "value": "10,000"
        },
        {
          "label": "Actual duration",
          "value": "4 days"
        },
        {
          "label": "Manual estimate",
          "value": "10 engineering weeks"
        }
      ],
      "duration_days": 4,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "10,000 lines of Scala migrated to Java in four days, against an estimate of ten engineering weeks by hand",
          "metric": "4 days vs 10 engineering weeks",
          "quote": "a team used Claude models to migrate 10,000 lines of Scala to Java in four days, a project estimated at ten engineering weeks by hand"
        }
      ],
      "techniques": [
        "language migration",
        "unblocking a jdk upgrade"
      ],
      "caveats": "Vendor-published. No cost or team-size figure given.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "altisource-amazon-q-350k-lines-java",
      "title": "Altisource modernizes 350,000 lines of legacy Java with Amazon Q Developer",
      "url": "https://www.shippedbyagents.dev/stories/altisource-amazon-q-350k-lines-java/",
      "source_url": "https://aws.amazon.com/solutions/case-studies/altisource-case-study/",
      "source_name": "AWS case study",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Amazon Q Developer"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Legacy Java modernized",
          "value": "350,000+ lines"
        },
        {
          "label": "Developer productivity",
          "value": "+25% (one modernization team, 'up to')"
        },
        {
          "label": "Story points (3-month avg per engineer)",
          "value": "~160 → 200"
        },
        {
          "label": "New applications delivered",
          "value": "4 in 4 months"
        },
        {
          "label": "Code vulnerabilities",
          "value": "−54%"
        },
        {
          "label": "Prior feature lead time",
          "value": "9–12 months"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "More than 350,000 lines of legacy Java modernised",
          "metric": "350,000+ lines",
          "quote": "Developers modernized more than 350,000 lines of legacy Java code"
        },
        {
          "claim": "Four new applications delivered in four months, where new features used to take 9 to 12 months",
          "metric": "4 apps in 4 months",
          "quote": "delivered four new applications in just four months"
        },
        {
          "claim": "Up to 25% productivity boost on one modernisation team, with story points per engineer up from about 160 to 200 over three months",
          "metric": "+25%",
          "quote": "the three-month Scrum story point average for software engineers increased from around 160 to 200 points"
        },
        {
          "claim": "Code vulnerabilities down 54%",
          "metric": "-54%",
          "quote": "Code vulnerabilities have decreased by 54 percent"
        }
      ],
      "techniques": [
        "ai-driven development lifecycle",
        "mcp servers for jira, bitbucket and figma"
      ],
      "caveats": "Vendor-published. The 25% figure is described as 'up to' and for one modernization team; story points are a self-reported velocity measure.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "dtcc-amazon-q-developer-poc",
      "title": "DTCC reports 40% higher developer throughput in an Amazon Q Developer proof of concept",
      "url": "https://www.shippedbyagents.dev/stories/dtcc-amazon-q-developer-poc/",
      "source_url": "https://aws.amazon.com/solutions/case-studies/dtcc-case-study/",
      "source_name": "AWS case study",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Amazon Q Developer"
      ],
      "models": [],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Developer throughput (POC)",
          "value": "+40% average"
        },
        {
          "label": "Code defects (POC)",
          "value": "−30% average"
        },
        {
          "label": "Security scores across repos",
          "value": "+5%"
        },
        {
          "label": "Example task time",
          "value": "10 hours → 6 hours"
        },
        {
          "label": "POC length",
          "value": "4 months (2024)"
        },
        {
          "label": "GenAI oversight team",
          "value": "40 people"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Developer throughput up 40% during the proof of concept",
          "metric": "+40%",
          "quote": "During the POC, the company boosted developer throughput by 40 percent"
        },
        {
          "claim": "A task that used to take 10 hours took 6",
          "metric": "10 h → 6 h",
          "quote": "something that used to take 10 hours now takes 6 hours using Amazon Q Developer"
        },
        {
          "claim": "Code defects down 30% while test coverage improved",
          "metric": "-30% defects",
          "quote": "reduced code defects and the associated operational risks by 30 percent while improving test coverage"
        },
        {
          "claim": "Security scores across code repositories up 5%",
          "metric": "+5%",
          "quote": "improved security scores across its code repositories by 5 percent"
        }
      ],
      "techniques": [
        "proof of concept before rollout",
        "phased rollout",
        "genai oversight team"
      ],
      "caveats": "Vendor-published; results are from the proof-of-concept phase. Build-failure rates reportedly did not increase.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "palo-alto-networks-cody-claude-2000-devs",
      "title": "Palo Alto Networks onboards 2,000 developers to a Sourcegraph Cody + Claude coding assistant in 3 months",
      "url": "https://www.shippedbyagents.dev/stories/palo-alto-networks-cody-claude-2000-devs/",
      "source_url": "https://aws.amazon.com/partners/success/palo-alto-networks-anthropic-sourcegraph/",
      "source_name": "AWS Partner case study",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Sourcegraph Cody"
      ],
      "models": [
        "Claude (via Amazon Bedrock)"
      ],
      "work_type": "Company rollout",
      "codebase": "n/a",
      "key_numbers": [
        {
          "label": "Developers onboarded",
          "value": "2,000 in 3 months"
        },
        {
          "label": "Pilot group",
          "value": "150 developers"
        },
        {
          "label": "Coding productivity",
          "value": "Up to +40%, average 20–25%"
        },
        {
          "label": "Total developers",
          "value": "3,500 worldwide"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "2,000 developers onboarded three months after the initial architecture",
          "metric": "2,000 in 3 months",
          "quote": "the company onboarded 2,000 developers to its generative AI coding tool"
        },
        {
          "claim": "Coding productivity up as much as 40%, with a company-wide average of 20 to 25%",
          "metric": "+20–25% average",
          "quote": "leadership saw coding productivity increase up to 40 percent, with the company-wide average between 20 and 25 percent"
        }
      ],
      "techniques": [
        "self-hosted single-tenant deployment",
        "pilot group before wider rollout"
      ],
      "caveats": "Vendor-published (AWS with partners Anthropic and Sourcegraph). Productivity measurement method not described.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "wiz-claude-code-python-to-go-pdf-library-20-hours",
      "title": "Wiz ports a 50,000-line Python PDF library to Go in about 20 hours with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/wiz-claude-code-python-to-go-pdf-library-20-hours/",
      "source_url": "https://claude.com/customers/wiz",
      "source_name": "Claude customer story (Anthropic)",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Language port",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Source size",
          "value": "~50,000 lines of Python (pypdf)"
        },
        {
          "label": "Output size",
          "value": "18,413 lines of Go"
        },
        {
          "label": "Time",
          "value": "~20 hours (10 active development, 10 testing)"
        },
        {
          "label": "Estimate without AI",
          "value": "2–3 months of specialized engineering"
        },
        {
          "label": "Test cases passed",
          "value": "500 of 500 pathological PDFs"
        },
        {
          "label": "Production result",
          "value": "2x+ faster PDF processing"
        },
        {
          "label": "Second port",
          "value": "20,000-line C++ library to Go in 2 days"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "A 50,000-line Python library ported to 18,413 lines of Go in about 20 hours, where Wiz estimated two to three months of a specialised engineer's time by hand",
          "metric": "~20 hours vs 2–3 months",
          "quote": "Claude Code migrated the 50,000-line Python library into an 18,413-line Go codebase in roughly 20 hours"
        },
        {
          "claim": "The new library processes PDFs at least 2x faster and removed the need for sandboxing",
          "metric": "2x+",
          "quote": "now processes PDFs at least 2x faster than the original Python implementation"
        },
        {
          "claim": "A second port, 20,000 lines of C++ to Go, took two days",
          "metric": "2 days",
          "quote": "converting fastText, a 20,000-line C++ data classification library, to Go in just two days"
        },
        {
          "claim": "Wiz's top 100 contributors saw a 1.5x increase in merged PRs since adopting AI coding tools",
          "metric": "1.5x"
        }
      ],
      "techniques": [
        "pathological test cases",
        "feature flag rollout",
        "debugging against real customer data"
      ],
      "caveats": "Published by Anthropic as a customer story. The 2–3 month estimate and the '20-30x estimated output multiplication' are the company's own estimates. The page shows no publish date.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "lg-cns-claude-code-build-factory-2888-apis-1340-screens-seven-months",
      "title": "LG CNS Build Factory migrates 2,888 of 2,913 APIs and 1,340 screens from a 20-year-old system in seven months with Claude Code",
      "url": "https://www.shippedbyagents.dev/stories/lg-cns-claude-code-build-factory-2888-apis-1340-screens-seven-months/",
      "source_url": "https://claude.com/customers/lg-cns",
      "source_name": "Anthropic customer story (LG CNS / Hyosup Bae)",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Claude Code"
      ],
      "models": [
        "Claude Opus 4.6"
      ],
      "work_type": "Migration",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "API conversion",
          "value": "2,888 of 2,913 APIs (99.1%)"
        },
        {
          "label": "Frontend screens",
          "value": "1,340 MiPlatform → React"
        },
        {
          "label": "Calendar",
          "value": "Nov 2025 – Jun 2026 (7 months), system went live June 2026"
        },
        {
          "label": "Cost vs conventional rebuild",
          "value": "Roughly 50%"
        },
        {
          "label": "Build Center",
          "value": "Claude Code standard across 200 engineers"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "2,888 of 2,913 APIs converted from a 20-year-old legacy system (99.1%)",
          "metric": "99.1% (2,888/2,913)",
          "quote": "Converted 2,888 of 2,913 APIs from a 20-year-old legacy system, a 99.1% completion rate"
        },
        {
          "claim": "Multi-million-dollar migration delivered in seven months at roughly half conventional rebuild cost",
          "metric": "~50% of conventional rebuild cost",
          "quote": "Delivered a multi-million dollar system migration in seven months, at roughly 50% of the cost of a conventional rebuild"
        },
        {
          "claim": "1,340 screens migrated from MiPlatform to React with parallel backend/frontend tracks",
          "metric": "1,340 screens",
          "quote": "Migrated 1,340 screens from a proprietary legacy frontend platform to React"
        }
      ],
      "techniques": [
        "yaml-scoped task decomposition",
        "file-chained context handoffs",
        "thin multi-layer harness",
        "24/7 parallel backend frontend tracks",
        "amazon bedrock deployment"
      ],
      "caveats": "Anthropic customer marketing page with no publish date. Cost comparison is approximate (“roughly 50%”) with no absolute figures. Customer system details are anonymized beyond sector. Multi-million / multi-billion-won claims are vendor-framed.",
      "updated": null,
      "update_note": null
    },
    {
      "slug": "satispay-claude-code-java-8-to-21-core-payment-under-four-days",
      "title": "Satispay modernizes its core payment service from Java 8 to Java 21 and Spring in under four days with Claude Code, against a four-week estimate",
      "url": "https://www.shippedbyagents.dev/stories/satispay-claude-code-java-8-to-21-core-payment-under-four-days/",
      "source_url": "https://claude.com/customers/satispay",
      "source_name": "Anthropic customer story (Satispay / Fabio Rapposelli)",
      "source_type": "vendor",
      "date": null,
      "tools": [
        "Claude Code"
      ],
      "models": [],
      "work_type": "Upgrade / modernisation",
      "codebase": "brownfield",
      "key_numbers": [
        {
          "label": "Core payment modernization",
          "value": "Under 4 days vs 4-week estimate (Java 8 → 21 + Spring)"
        },
        {
          "label": "Code from Claude",
          "value": ">75% of code committed each month"
        },
        {
          "label": "Adoption",
          "value": ">90% of engineers; full laptop rollout in 30 days"
        },
        {
          "label": "Roadmap compression",
          "value": "18-month update goals done in 7 months"
        },
        {
          "label": "Lambda transforms",
          "value": "3–5 days → under 1 hour"
        }
      ],
      "duration_days": null,
      "cost_usd": null,
      "prs_merged": null,
      "claimed_benefits": [
        {
          "claim": "Core transaction service Java 8 → Java 21 and Spring upgrade in under four days vs four-week estimate",
          "metric": "<4 days vs 4 weeks (10x)",
          "quote": "safely completing a Java 8 to Java 21 and major Spring upgrade in under four days against a four-week estimate"
        },
        {
          "claim": "Over 75% of code committed each month generated with Claude",
          "metric": "75%+",
          "quote": "Generates over 75% of code committed to Satispay’s repositories each month"
        },
        {
          "claim": "Data-transform Lambdas cut from three to five days to under an hour",
          "metric": "3–5 days → <1 hour",
          "quote": "Lambdas would routinely take over three days to build. Now it takes less than an hour."
        }
      ],
      "techniques": [
        "managed device fleet rollout",
        "claude review before human review",
        "encoded java scaffolding",
        "modernization subagents",
        "mcp-connected lambda generator"
      ],
      "caveats": "Anthropic customer marketing page. No publish date on the page. No independent verification of the four-day migration, defect rates, or the 75% code-share metric. CTO quotes selected by the vendor.",
      "updated": null,
      "update_note": null
    }
  ]
}