Cockroach Labs engineers describe a five-month experiment that structured AI coding agents around a teaching-hospital metaphor, with roles such as Triage Nurses, Fellows, Review Attendings and Discharge Nurses enforcing mandatory planning, tests and second-opinion reviews before anything merged into MOLT, the company’s database migration tool set. The pipeline ran on GitHub Actions, with the state of each ‘patient’ issue held in labels, and processed more than a million lines of code with only seven reverts. The authors say IBM Db2 support that would have taken about nine months and some $160,000 of engineering time, the way Oracle support did in 2024, landed in under two days for roughly $4,172 in tokens with no human writing the code, and argue the deliberately slow review process is what made agent-written code trustworthy enough for a database product.
Five months treating bugs like patients and coding agents like a medical team
Cockroach Labs engineers describe a five-month experiment that structured AI coding agents around a teaching-hospital metaphor, with roles such as Triage Nurses, Fellows, Review Attendings and Discharge Nurses enforcing mandatory planning, tests and second-opinion reviews before anything merged into MOLT, the company's database migration tool set. The pipeline ran on GitHub Actions, with the state of each 'patient' issue held in labels, and processed more than a million lines of code with only seven reverts. The authors say IBM Db2 support that would have taken about nine months and some $160,000 of engineering time, the way Oracle support did in 2024, landed in under two days for roughly $4,172 in tokens with no human writing the code, and argue the deliberately slow review process is what made agent-written code trustworthy enough for a database product.
Source: Cockroachlabs