Human review: 38 sourced coding agent stories
A person reviews or approves the change (or the plan) before merge, or the owning team makes the final call. This technique sits in the Review and merge stage. Of the 38 stories, 12 were published by a tool vendor and 26 by the people who did the work. 28 involve an existing codebase and 9 started from scratch. Every number links to where it was published.
Keboola: a two-sentence prompt to Claude Code merges 13 Zend-removal PRs in about 22 hours through an existing agent harness
Keboola · Claude Code
CaveatsWritten 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.
13 in about 22 hours (3 more open)
Merged PRs
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
University of Tokyo and NIES (NICAM) · Claude Code
CaveatsWritten 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.
~140,000 lines of Fortran analysed and ported into ~35,000 lines of C++/Kokkos
Code ported
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
Tabby (taewoo) · Claude Code
CaveatsSelf-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.
19 days since 5 Oct; v1.2 about a day from plan to release
Time
A database specialist who doesn't read Rust ships ShuttleSCP, a 22,000-line Rust and GTK 4 file-transfer client, with Claude Code
ShuttleSCP (David, dsalvado) · Claude Code
CaveatsSelf-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.
About 22,000 lines of Rust, 149 automated tests
Size
Repo fixer: a scheduled CircleCI pipeline with LLM fixers opens 150 migration PRs across an org overnight and keeps fixing new repos
Claire Alvis (Confident Commit) · repo fixer
CaveatsSelf-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.
150 open PRs across the org by the next day
Migration PRs
Uber's AutoClean agents remove stale feature flags across its monorepos, with over 16,000 flags and 1,000,000 lines cleaned
Uber · AutoClean (Uber), PolyglotPiranha (Uber, open source), CodeCourt (Uber)
CaveatsUber'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.
"more than 16,000 flags" ("Over the past few years")
Flags cleaned
Kassebaum Engineering lets agents review and merge routine code: 136 PRs across four repos in ten days, with people merging the riskiest tier
Kassebaum Engineering (William R. Kassebaum) · Claude Code, Cursor, Grok CLI
CaveatsSmall 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.
136 across four repositories over ten days in October 2026
Merged PRs
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
Anthropic (Claude Code team) · Claude Code
CaveatsVendor 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.
claude.dev blog (Addy Osmani, Morgan Westlee Lunt, Qing Wang)
More than doubled from June to September
Merged PRs
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
Meta · Tool unstated
CaveatsMeta'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.
Core feature parity with Facebook Groups in eight weeks
Time to feature parity
Foxbox Digital ships a production marketing site with two engineers in 14 days using Claude Code
Foxbox Digital (agency) · Claude Code
CaveatsWritten by the agency that did the work, partly as marketing for its services. No cost, token usage or pre-AI estimate is given.
2 weeks
Calendar time
Gemini 4 Argon runs unattended for 11 hours on the 205,000-line ADK Python codebase; 2 of its 6 optimizations ship
Google · Google Antigravity
CaveatsVendor 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.
~205,000 lines, ~800 source files, ~17,000 unit tests
Codebase
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
Cockroach Labs · Claude, MOLT Sinai (Cockroach Labs)
CaveatsSelf-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.
"less than two days" vs 9 months for Oracle support in 2024
Db2 support time
Block uses Claude Fable to orchestrate code migrations of thousands of pull requests, with two humans approving every deploy
Block · Buzz (Block)
CaveatsThis 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.
"hundreds or thousands of individual pull requests across multiple repos and millions of lines of code"
Migration size
Juno merges 26 of its last 30 PostHog self-driving PRs, a roughly 90% merge rate, with the CEO still reviewing each one
Juno · PostHog self-driving, PostHog scout
CaveatsVendor-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.
Roughly 90%
Self-driving PR merge rate
Comarch compresses a five-year rebuild of its 10-million-line Clarion ERP to 18 months with Devin
Comarch · Devin
CaveatsPublished 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.
About 10 million lines of Clarion, 20 years old
Codebase
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
Mistral AI (for an unnamed European energy operator) · Mistral Vibe
CaveatsVendor 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.
40,000 of 300,000 lines of Fortran 77 (first sprint, core functionality)
Lines migrated
DoorDash's multi-agent system turns stale feature flags into merge-ready PRs for $4.79 and 14 minutes each
DoorDash · Claude
CaveatsSelf-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.
45 of 50 stale flags
Usable PRs
Asana finishes a five-year Enzyme → React Testing Library migration in two weeks with Codex
Asana · OpenAI Codex
CaveatsPublished 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.
OpenAI customer story (Asana also published its own account, in partnership with OpenAI)
~2 weeks
Calendar time
Vercel's AI SDK software factory authors 25–35% of merged PRs and closes most issues, four weeks into production
Vercel · ai-sdk-factory (Vercel)
CaveatsSelf-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.
25–35%
Share of weekly merged PRs authored by the factory
Wix's Code Migration Platform has a coding agent open migration PRs across hundreds of repos, saving 650 developer-days in a quarter
Wix · Code Migration Platform
CaveatsSelf-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.
650 in a single quarter (their estimate)
Developer-days saved
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
ORB5X (PSI, EPFL, Max-Planck-Institut für Plasmaphysik) · Cursor, Claude Code, OpenCode, OpenAI Codex
CaveatsWritten 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.
Almost machine accuracy for ITG, ITPA and chirping field components
Benchmark agreement with Fortran ORB5
Solo staff engineer builds a 325,000-line workflow automation platform in 11 months of evenings and weekends with Claude Code
Flywheel (Travis Wilson) · Claude Code
CaveatsSelf-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)".
325,000 lines of production code (170K Go, 145K TypeScript, 7.5K Python, 2,500 Terraform)
Codebase
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
Bloomberg (with University College London) · Pomona
CaveatsIndustrial 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.
39 over 55 workdays; 32 merged (82.1%)
PRs
Eventship replaces a quoted 6-month, 5-person rebuild with 5 weeks of solo Claude Code work and migrates 20,000 users
Eventship · Claude Code
CaveatsSelf-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.
5 weeks, solo
Time
Atlassian's Rovo Dev code review agent cuts median PR cycle time 30.8% across 1,900+ internal repositories in a year-long evaluation
Atlassian · Rovo Dev
CaveatsAtlassian 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.
Over a year across 1,900+ repositories
Scope
Ten months of GitHub's Copilot coding agent in dotnet/runtime: 878 PRs, 535 merged, with a revert rate no worse than humans'
Microsoft (.NET team) · GitHub Copilot
CaveatsMicrosoft 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.
878; 535 merged, 253 closed, 90 open (67.9% success)
CCA PRs in dotnet/runtime
Taller Technologies’ Chiron platform cut three modernization programs from 36 to 9.3 summed project-weeks, with gains coming only once agents were orchestrated
Taller Technologies · Chiron
CaveatsBoth 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.
arXiv (Maximiliano Armesto and Christophe Kolb, Taller Technologies)
36.0 weeks (traditional) to 9.3 weeks (V4), 3.87x faster
Summed project duration
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
SoundSignal (W. David Betts) · Claude Code
CaveatsSelf-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.
5 calendar days
Time
Stripe's unattended Minions agents produce over 1,300 merged pull requests a week, human-reviewed but with no human-written code
Stripe · Goose, Minions, Devbox, Blueprints
CaveatsSelf-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.
Over 1,300 a week, completely minion-produced (up from 1,000 at Part 1)
Merged PRs
Fullscript's Nitro background agent creates 19% of pull requests three months after launch on a decade-old Rails monolith
Fullscript · Claude Code, Nitro
CaveatsSelf-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.
19% in January 2026, three months after launch
Share of PRs created by Nitro
Solo developer builds a 200,000-line multi-tenant SaaS with four portals and four APIs in three months with Claude Code
Saru (ko-chan) · Claude Code
CaveatsSelf-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.
~200K (Go 84K + TypeScript 113K)
Lines of code
Vellum's 15 engineers double monthly PRs from 946 to 1,910 with background agents and Codex as first-pass reviewer
Vellum · Devin, OpenAI Codex, Cursor, Claude Code
CaveatsSelf-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.
946 in May → 1,910 in October
PRs
Mitchell Hashimoto ships a non-trivial Ghostty feature with Amp in 16 sessions for $15.98
Ghostty (Mitchell Hashimoto) · Amp
CaveatsA 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.
16
Agent sessions
Cloudflare's Kenton Varda prompts an OAuth 2.1 provider library out of Claude Code in about five days
Cloudflare · Claude Code
CaveatsThe 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.
~5 days of heavy work (with other duties), plus minor commits over the month
Time with AI
Google Ads uses LLMs to migrate 32-bit IDs to 64-bit in a 500M+ line codebase
Google · Google internal migration toolkit (fine-tuned Gemini)
CaveatsGoogle uses its own model; authors call it an experience report, not a controlled study.
500M+ lines
Codebase size
Google migrates 5,359 test files from JUnit3 to JUnit4 in 3 months with LLMs
Google · Google internal migration toolkit (fine-tuned Gemini)
CaveatsExperience report; Google uses its own model.
5,359
Files migrated
Nubank uses fine-tuned Devin to split a 6M-line ETL monolith, claiming 8–12x engineering-hour efficiency
Nubank · Devin
CaveatsVendor-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.
6M+ lines, ~100,000 data-class implementations
Codebase
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
Satispay · Claude Code
CaveatsAnthropic 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.
Under 4 days vs 4-week estimate (Java 8 → 21 + Spring)
Core payment modernization