A million lines of code (tests included). 1,238 pull requests. Seven reverts.

Cockroach Labs got those numbers from coding agents run like a teaching hospital. No code before a plan, no plan without review. A separate review agent checks every plan and every change, and nobody gets to weaken the tests to make something pass. A human has approved every merge.

My favourite detail is the final check. They call it the Discharge Nurse, and it doesn't re-review the code at all. It verifies that the review happened properly.

That flips how most of us think about agent quality. When agents write more than anyone can read, you stop asking whether every line is right and start asking whether every step actually happened. Plans, reviews and honest tests become the thing you inspect.

The cost story is striking too, with caveats. A migration to IBM Db2 took under two days for about $4,172 in tokens, against roughly nine months for their 2024 Oracle migration. That's token cost only, it's a best-case comparison, and the Db2 code is still being verified. Not production-proven yet.

The shape is still worth copying. Gate the process, then let the agents run.