Claude writes 80% of Anthropic’s code, and CI buckled

Three quick fixes to Anthropic’s test selection service bought 70 days, then 29 days, then less than a day. CI jobs grew 25x over six months, and Claude now authors 80% of the code engineers there ship.

The service picks which tests run on each pull request, and to do that it records results from every CI run. That recorder lived as a single process holding history in memory, so when it lagged, the selector chose tests off stale data.

The rebuilt version writes into an in-memory data store: any worker appends a result to a journal and moves on, which makes the recorder shardable. Anthropic’s write-up calls the distributed setup more expensive to run. It took one engineer three weeks; a year ago, closer to a quarter.**