Cockroach Labs puts review gates around its coding agents
Cockroach Labs reports adding Db2 migration support in under two days for $4,172 in model tokens. Its earlier Oracle work took nine months and about $160,000 in engineering time—an uncontrolled comparison, not a benchmark. The agent workflow reviews plans before implementation and checks code before merging. After the first autonomous week, the team required human approval for every merge.
- Plan + review
- Build + test
- Code review
- Human approval
Cockroach Labs · Oct 9, 2026
Why it matters
The useful lesson is how to make agent work inspectable, including when the process becomes too expensive or bureaucratic.
More context
Try the individual practices before copying the whole system: reviewed plans, structured handoffs and tests that fail without the fix. This is an internal five-month case study, not a publicly verified turnkey service. The authors report over $135,000 in model costs and describe limits on repeated review cycles.
Keep in mind
The Db2 experiment was still undergoing correctness confirmation when published. The Oracle comparison involves different work and cost categories; it is not a controlled benchmark.
Original read