Run coding agents like a medical team to fix bugs
Over five months, the author treated bugs like patients and organized coding agents as a hospital-style team to diagnose and fix them.
Done with a custom-built or unnamed agent
- What they did
- Cockroach Labs built a GitHub Actions pipeline in which labels on an issue trigger role-specific agents. A Fellow reproduces the bug and posts a treatment plan, and a separate Review Attending approves or rejects it before any code is written. Large issues are split into sub-issues of under 1,000 lines each. A Discharge Nurse audits the review before merge, and humans handle escalations.
- How it went
- Over five months it merged 1,238 PRs and over a million lines of code with 7 reverts. A Db2 migration feature shipped in under two days. Reviewers sometimes looped on trivial fixes, and agents filed about half of all issues.
- Worth knowing
- It used about $135,000 in Claude tokens, around $84 per issue. Some issues went 9 to 11 rework rounds before a circuit-breaker was added.
Try it yourself with Claude Code
Treat the bugs in this repo like patients in a hospital. For each bug in [issue list or file], have one agent triage and reproduce it, another diagnose the root cause, and another write a fix with a test, then have a separate reviewer check the fix. Finish with a summary table of each bug, its diagnosis, the fix and test status, and pause for my approval before committing or pushing anything.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on HN · Oct 9, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Edit site design visually and have an agent write the code
CSS Studio lets designers tweak a site visually in the browser and streams the changes via MCP to an existing coding agent that implements them in the codebase.
Test Devin on pulling Notion data into Sheets
Answer.AI testers found Devin completed only 3 of 20 tasks, succeeding at moving Notion data to Google Sheets and building a planet tracker.
Fix a broken Delete button in a GitHub repo
The author had Jules clone a repo and fix a Delete button so it disables deletion of content older than two hours, with tests added unprompted; the author reviewed and published the branch.
Build a homepage from Figma designs via MCP
The author connected Figma to Cursor over MCP to build a new homepage for a font archive, with great results after Lovable disappointed.
Build a WordPress marketing site from Figma overnight
bnistordev connected Devin to a Figma file via MCP and had it build an entire WordPress marketing site overnight.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.