saved
Five months treating bugs like patients and coding agents like a medical team
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Cockroach Labs engineers Adam Storm and Rafi Shamim describe MOLT Sinai, a GitHub Actions pipeline that models coding agents as a teaching hospital, with triage, reviewed treatment plans, I-PASS handoffs, discharge audits, and a human Chief. Over five months it merged 1,238 PRs for about $135,000 in Claude tokens and built IBM Db2 migration support in under two days, though it also proved bureaucratic, prolific, and poor at teaching junior engineers.
ideas
- Quality over throughput. Unlike speed-first factories such as Gas Town or Cursor’s SQLite rebuild, the hospital asks how many PRs they would be embarrassed to merge.
- Hospital rules as gates. No code before a reviewed plan, no improvising, no weakening tests, structured I-PASS escalation instead of flailing, a Discharge Nurse that audits the review, and a precedent log before bothering the human Chief.
- The numbers. From April to September: 1,238 PRs merged, 7 reverted, 92 issues reaching a human, about $84 per issue, and 85% of the Sinai framework’s own commits written by the pipeline.
- Personas and prompt rot. Role framing made agents behave like clinicians, but an audit found 23% of 100,000 words of skill files removable, costing tokens on every run.
- Costs of bureaucracy. One-line fixes looped through up to eleven rework rounds until a circuit-breaker landed, the Research Department flooded the backlog, and juniors get little learning from it.
quotes
“Db2 took less than two days, and no human wrote any of the code.”
“Speed matters, but quality matters much much more.”
“When you tell an LLM a persona to embody, it does it very very well.”
“Our prompts rotted the way code does: we kept adding rules and never removing them”