Five months treating bugs like patients and coding agents like a medical team

Why we picked it · the editor's summary

Cockroach Labs' MOLT team modeled its issue pipeline on a teaching hospital: triage agents, fellows that write fixes, a separate attending that reviews plans and code, and humans as chiefs of medicine. The rules do most of the work: no code before an approved plan, no weakened tests, a structured handoff instead of a retry, and an append-only log of past human rulings the agents check before escalating. The authors report about $135,000 in tokens over five months, roughly $84 per issue, and Db2 support built in under two days, though that code predates human approval and is still being verified. The process also overcorrected: an urgent one-line fix took eleven rework rounds before a circuit breaker went in. The figures are self-reported; the rule set and the failure list are what to copy.