Automated verification: 83 sourced coding agent stories
The agent's work is checked automatically: tests, compilers, linters, CI gates or fix loops, including parity harnesses, coverage or eval sets built before the change. This technique sits in the Verification stage. Of the 83 stories, 15 were published by a tool vendor, 67 by the people who did the work and 1 by independent researchers or press. 55 involve an existing codebase and 25 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
Christopher Weir coordinates about 90 Claude Code worker sessions from one "ringleader" thread for more than 60 merged PRs in five weeks
Christopher Weir · Claude Code
CaveatsSelf-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.
About 90 Claude Code sessions
Worker sessions
QuEra: one engineer with Cursor and Claude Code merges 111 PRs across 19 repos, empty repo to merged service in six days
QuEra · Cursor, Claude Code, Quantopus (QuEra)
CaveatsSelf-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.
111 across 19 repositories
Merged pull requests
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
Alexander Marquardt (e-commerce search solution) · Claude Code
CaveatsSelf-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.
1,652 (about 373 a week), median 492 changed lines
Merged PRs, main repo, October 2026
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
Operum · Operum
CaveatsWritten 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.
296 in 12 days (15–26 Oct 2026)
Merged PRs
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
Dropbox · Tool unstated
CaveatsWritten 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.
Dropbox Tech blog (Pravi Garg, Sourabh Yerfule, Gennadiy Civil)
22 across Home, All Files, file previews and Photos
Accessibility features
Duolingo builds AI STR Fix on the Claude Agent SDK to investigate and fix shake-to-report bugs, including release blockers
Duolingo · Claude Agent SDK
CaveatsWritten 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.
Has fixed release-blocking bugs
Outcome
Bloomberg's Automated Uplift ran LLM-assisted migrations across thousands of repositories: 2,333 PRs and 729 successful migrations
Bloomberg · Tool unstated
CaveatsWritten 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.
2,333 generated across two campaigns
Pull requests
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
Microsoft (Copilot Cowork) · Tool unstated
CaveatsSelf-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.
13 Jan to 16 Jun 2026 (~4.5 months)
Empty repo to worldwide GA
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
X2 Labs founder rebuilds BZFlag as a multiplayer browser game with Claude Code in seven days, mostly overnight, with 2,421 tests
X2 Labs (George Purusniuc) · Claude Code
CaveatsSelf-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.
7 days from first commit to public
Time
MerciYanis rewrites its whole platform, about 250,000 lines across 10 repositories, in one day with one engineer and around 30 parallel agents
MerciYanis · Tool unstated
CaveatsWritten 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.
Roughly 250,000 lines across 10 repositories
Scope
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
MyLocker (Michael Cutler) · Claude Code, OpenAI Codex
CaveatsSelf-reported by the builder. Analytics are off, so there are no user numbers. Unofficial app for one gym chain's members.
3 days
Backend cutover
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
Microsoft / UC Irvine / OpenAI researchers (while at Microsoft) · GitHub Copilot
CaveatsResearch 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.
9 repos, 3 language pairs
Subjects
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
Solo developer rebuilds his Astro site on Next.js 16 with Claude Code in two days: 228 posts reworked, 543 tests, no broken URLs
Alex Cloudstar (solo developer) · Claude Code
CaveatsAuthor'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.
Alex Cloudstar (DEV Community, originally on alexcloudstar.com)
Two days (late 21 Oct to morning of 23 Oct 2026)
Time
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
ParkEase (Onkar Deokate) · Claude Code
CaveatsSelf-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.
Five weeks (Sep 10 – Oct 13, 2026 per the repo)
Time
GetYourGuide's Automerge merges 5,000+ dependency PRs a month across 250+ repos, with an AI agent fixing broken CI
GetYourGuide · Automerge (GetYourGuide)
CaveatsSelf-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.
"more than 5,000" a month with "zero human clicks"
Dependency PRs merged
Solo developer ships Kempt, a signed and notarised macOS app with Stripe licensing, four days after the first commit, built mostly by Claude Code
Kempt (Yassir Halaoui) · Claude Code, OpenAI Codex
CaveatsSelf-reported by a consultant whose post ends with a contact address. No cost figures. No sales or user numbers yet.
First commit 13 Oct to 1.0 on 17 Oct 2026 (five calendar days)
Time
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
Keepard (Rodion, personal project) · OpenAI Codex, Claude Code
CaveatsSelf-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.
Nine calendar days, about 3 to 4 hours of attention a day
Build time
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
LokalBot (Stevan Bogosavljević) · Claude Code, OpenAI Codex
CaveatsSelf-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.
About 150,000 lines of Swift
Source app size
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
Three people build Loft with coding agents over three months, then ship its open format 53 hours after the first commit
Loft · Claude Code, OpenAI Codex
CaveatsWritten 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.
3 people
Team
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
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
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
Lamplit Labs (Manish Tiwari) · Tool unstated
CaveatsSelf-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.
41k-line .NET Framework 4.8 service
Codebase
Claude Code ships 19 minor releases of the Smarter Django platform in 18 days, adding about 238,000 lines
Smarter · Claude Code
CaveatsSelf-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.
19 minor and 45 patch releases in 18 days
Releases
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
Solo developer ships Scumble, a desktop image editor in the Microsoft Store, five weeks after the first commit with Claude Code
Scumble (Dennis Schöneberg) · Claude Code
CaveatsSelf-reported by the developer. No cost figures. Line counts are approximate.
Five weeks from first commit to release 43
Time
Lucas Firmo's agent fleet moves a production iOS app to Swift 6 strict concurrency in 10 days, with zero warnings left
· Claude Code
CaveatsA 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.
10 days for the Swift 6 strict concurrency migration
Time
An autonomous Claude Code agent ships 340+ production deploys behind a three-stage gate it cannot edit
Yureki Lab (dev.to author) · Claude Code
CaveatsSelf-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.
"340+ production deploys with zero human-paged incidents" over six months
Deploys since the gate
Token Clicker: a 33,716-line browser idle game shipped 19 hours after its specs with Claude Code
Julien Renaux (independent developer) · Claude Code
CaveatsThe 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.
19 hours 39 minutes wall-clock (night included)
Specs to launch
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
GitHub ports the Copilot agent runtime from TypeScript to 800K+ lines of Rust, mostly by one engineer
GitHub (Microsoft) · GitHub Copilot
CaveatsPublished 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.
832,378 lines
Production Rust written
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
Solo builder ships SkillPilot, a paid product, with Claude Code in about three weeks, but 1,805 passing tests missed a dead checkout
SkillPilot (Carlos Infantes) · Claude Code
CaveatsSelf-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.
About three weeks
Build time
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
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
Assembled auto-approves low-risk PRs with Codex and Claude Code reviewers, lifting PRs merged to 2.4× its pre-agent baseline
Assembled · OpenAI Codex, Claude Code, Devin, 143.dev (Assembled)
CaveatsSelf-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.
2.4× pre-agent baseline
PRs merged
Ambrook migrates its design system from Material UI to Base UI in just over two weeks with a graph of planner and worker agents
Ambrook · Claude Code, Niteshift
CaveatsWritten 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.
54,000 lines changed across 794 files
Change
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
Anthropic's Mike Krieger ports an internal Python tool to 165,000 lines of TypeScript over a weekend with hundreds of Claude Code agents
Anthropic (Mike Krieger) · Claude Code
CaveatsVendor-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.
165,000 lines of TypeScript over a weekend
Output
Claude plans and reviews, Codex builds: a production music site migrated from a React SPA to Next.js in 36 hours
Nik McFly · Claude, OpenAI Codex
CaveatsSelf-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.
36 hours
Migration time
Bun's 535,000-line Zig codebase is ported to Rust in 11 days by up to 64 parallel Claude Code agents
Bun (Oven / Anthropic) · Claude Code
CaveatsBun 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).
535,496 lines of Zig (excluding comments), 1,448 .zig files
Source size
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
Gusto · Claude Code
CaveatsA 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.
How I AI podcast (Claire Vo) with Eddie Kim, Lenny's Newsletter
Four engineers (including the CTO) and one designer
Team
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
Datadog migrates its live Stream Router from FoundationDB to PostgreSQL with Claude and Cursor; 45-minute operations now take ~1 second
Datadog · Claude, Cursor
CaveatsSelf-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.
~3 months (Dec 2025 – Feb 2026); proof of concept in 4 weeks
Design to production
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
Alfred Wegener Institute (FESOM2) · Claude Code
CaveatsWritten 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.
73,991 lines in 109 files (core ocean–sea-ice configuration)
Fortran ported
Sixt's custom Claude agent migrates 780 iOS test files (~7,000 tests, 108,914 lines) from XCTest to Swift Testing for about $106
Sixt · Claude Agent SDK, Swift Testing migration agent (Sixt)
CaveatsWritten 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.
780 across 50 modules
Files processed
Dropbox’s internal coding agent platform Nova accounts for roughly 1 in 12 pull requests, and runs flaky-test fixes and migrations
Dropbox · Nova, Deflaker
CaveatsWritten 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.
Roughly 1 in 12 pull requests at Dropbox
Share of pull requests
Stripe builds continuous Java upgrades with Minions fixing CI failures, reaching 93%+ of services on JDK 17 within a week of rollout
Stripe · AutoJDK, Minions
CaveatsMuch 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.
93%+ of services within under a week of rollout
JDK 17 adoption
Plaid’s Hamilton cloud coding agent writes thousands of PRs and accounts for about 20% of all merges
Plaid · Hamilton
CaveatsSelf-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.
Roughly 20% of all merges
Hamilton share of merges
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
Baohua (Luke Lin) · Claude Code, OpenAI Codex, gstack
CaveatsSelf-reported by the builder. No user or revenue numbers. The post recommends tools by name (gstack, Venn AI).
April 1 to April 30, 2026
Time to ship
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
Google (AI and Infrastructure) · Tool unstated
CaveatsGoogle 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.
6.4x to 8x over manual migration
Speedup
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
Sergio Azócar · Claude Code
CaveatsPersonal 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.
~1,670 TypeScript files, 7 apps
Monorepo size
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
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
ZoomInfo · Claude Code
CaveatsSelf-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.
938 files
Scope
Vercel auto-merges 58% of PRs in its largest monorepo without human review, cutting average merge time from 29 hours to 10.9
Vercel · Vercel Agent, LLM PR risk classifier (Vercel), Vercel AI Gateway
CaveatsVendor-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.
58%
PRs merged without a human reviewer
Jmix migrates a CUBA Platform Timesheets app to Jmix with Codex in 25 hours against a 215-hour estimate
Jmix (Haulmont) · OpenAI Codex
CaveatsPublished 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.
25 man-hours
Actual effort
Eve’s Slack-driven background agent WALL-E authors 22% of merge requests in its monorepo, 53% of them one-shot
Eve (Eve Legal) · OpenCode, WALL-E
CaveatsWritten 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.
22% of all merge requests in the monorepo
Share of merge requests
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
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
Datadog · Claude Code, OpenAI Codex, pup
CaveatsCompany 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.
87% reduction after agent-guided optimisation (started at 8x Redis 8.4's memory)
redis-rust memory
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
Composio (Prateek) · Claude Code
CaveatsShort 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.
8 days
Build time
Cloudflare engineer rebuilds the Next.js API surface on Vite in under a week with Claude in OpenCode, for about $1,100 in tokens
Cloudflare · OpenCode
CaveatsWritten 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.
Under a week; first commit 13 February
Time
Ladybird ports its JavaScript engine's parser and bytecode compiler from C++ to Rust in two weeks with Claude Code and Codex
Ladybird · Claude Code, OpenAI Codex
CaveatsSelf-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.
About 25,000 lines of Rust
Output size
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
OpenAI team ships an internal product with zero hand-written code: ~1M lines and ~1,500 merged PRs in five months with Codex
OpenAI · OpenAI Codex
CaveatsWritten by the vendor about its own tool. Line count includes docs and tooling. No cost, time-vs-estimate or quality figures.
~5 months
Duration
16 parallel Claude agents build a 100,000-line Rust C compiler that builds Linux, for ~$20,000
Anthropic (Nicholas Carlini) · Claude Code
CaveatsVendor-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.
Just under $20,000
API cost
Instacart's Caper team migrates its smart-cart Android app to Jetpack Compose with AI, 5-7x faster on navigation
Instacart · Tool unstated
CaveatsNo 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.
5-7x
Phase 2 speed-up
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
Cisco uses Codex across 15+ repos to cut build times ~20% and speed up C/C++ defect fixing
Cisco · OpenAI Codex
CaveatsVendor-published.
1,500+ per month
Engineering hours saved
Simon Willison ports an HTML5 parser from Python to JavaScript with Codex CLI in about 4.5 hours
Simon Willison (independent developer) · OpenAI Codex
CaveatsStraight 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.
~4–4.5 hours
Idea to finished library
Emil Stenström builds JustHTML, a pure-Python HTML5 parser passing all 9,200+ html5lib tests, with Copilot agent mode
Emil Stenström (independent developer) · GitHub Copilot
CaveatsFact 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.
Simon Willison (simonwillison.net), summarising Emil Stenström's write-up
A couple of months of off-hours
Time
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
Lee Robinson migrates cursor.com off a headless CMS back to code and Markdown in three days with Cursor agents for $260
Cursor (Lee Robinson) · Cursor
CaveatsWritten 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).
3 days (estimated 1–2 weeks)
Calendar time
Claude Code ports a 25-year-old Linux tape driver from kernel 2.4 to 6.8 over two evenings
Dmitry Brant (independent developer) · Claude Code
CaveatsNeeded 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.
Two evenings
Time
Solo developer ships a 20,000-line native macOS app, writing fewer than 1,000 lines by hand
Indragie Karunaratne (independent developer) · Claude Code
CaveatsNo 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.
~20,000 lines
Codebase size
Airbnb migrates ~3.5K Enzyme test files to React Testing Library in 6 weeks with LLMs
Airbnb · Tool unstated
CaveatsModel names and dollar cost not disclosed. Published date is approximate (March 2025, per InfoQ coverage).
~3,500
Test files migrated
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 estimates ~89% time saved on Joda-Time → java.time migration clusters
Google · Google internal migration toolkit (fine-tuned Gemini)
CaveatsOngoing migration; estimate comes from tech leads comparing against the manual approach, extrapolated from a sample of changelists.
~89%
Time saved on small clusters (estimate)
Wiz ports a 50,000-line Python PDF library to Go in about 20 hours with Claude Code
Wiz · Claude Code
CaveatsPublished 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.
~50,000 lines of Python (pypdf)
Source size