Orchestrator and worker agents: 27 sourced coding agent stories
A lead, planner or orchestrator agent splits the work and hands pieces to worker or specialist agents, sometimes with a judge. This technique sits in the Doing the work stage. Of the 27 stories, 8 were published by a tool vendor, 18 by the people who did the work and 1 by independent researchers or press. 20 involve an existing codebase and 6 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
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
Ticksly, a three-person concert ticketing startup, has agents write all its code: 2,540 merge requests in 28 days at about $8 each
Ticksly · Claude Code, OpenAI Codex
CaveatsWritten 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.
2,540 opened in the last 28 days
Merge requests
Spawn, a ten-person game engine startup, says its agent fleet shipped 77,000 substantial changes in 30 days on $3.1M of tokens
Spawn · Claude Code
CaveatsThese 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).
X Article by Spawn's founder (@jsnnsa), "How I Built the Most AI-Pilled Company in the World"
77,000 substantial changes in the last 30 days, about 108 an hour
Changes shipped
DANA Indonesia cut a design-system component migration from 24 hours to 52 minutes per batch with Cursor agents run from Notion
DANA Indonesia · Cursor
CaveatsVendor 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.
52 minutes, down from an average of 24 hours
Time from assignment to review-ready PR per batch
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
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
Sparq builds a 22-agent internal platform with Claude Code: 34,000 lines across 453 commits
Sparq · Claude Code
CaveatsSparq 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.
~34,000 Python
Lines of code
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
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
GPXplore (Brian Erickson) · Claude Code, OpenAI Codex, Lovable
CaveatsSelf-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.
134 days from first prototype commits to the 1 October App Store launch
Calendar 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
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
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
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
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
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
National Australia Bank builds a greenfield Android payment app in under three weeks with Cursor, against a four-month scope
National Australia Bank (NAB) · Cursor
CaveatsPublished 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.
Under 3 weeks vs. 4-month original scope
Payment app
Cloudflare’s OpenCode-orchestrated AI code review runs 131k times across 5,169 repos in 30 days, median 3m 39s and $0.98
Cloudflare · OpenCode
CaveatsSelf-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.
131,246 review runs on 48,095 MRs across 5,169 repos
30-day volume
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
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
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
Cursor's agent swarm writes a 1M+ line browser in about a week, and critics find it barely compiles
Cursor (Wilson Lin) · Cursor
CaveatsVendor 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.
Close to a week
Duration
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
LG CNS · Claude Code
CaveatsAnthropic 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.
2,888 of 2,913 APIs (99.1%)
API conversion
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