Anthropic rebuilt its CI when Claude started writing most of the code

Anthropic's engineers now ship about eight times as much code per quarter as they did in the 2021 to 2025 baseline, and Claude writes roughly 80% of it. That volume, combined with a tenfold growth in tests and almost no new hiring, pushed continuous integration jobs up 25x in six months. The part that broke first was test impact analysis, the service that works out which tests a given change actually needs to run. It was a single process, and it could not keep up.

The team writes that they tried three quick fixes before rebuilding. A bigger machine with double the cores bought 70 days. Sharding the work per package bought 29 days. Daily restarts to clear memory lasted less than a day. Each patch delayed the moment when unprocessed results piled up faster than the service could clear them, which is dangerous because stale results feed back into decisions about later pull requests.

The redesign splits the work. Stateless listener workers append raw results to a journal, and a separate consumer rolls that journal into queryable history, so the listeners can scale horizontally. Building it took three weeks, down from a quarter for a comparable effort a year earlier, which the author reads as a sign of the same acceleration that caused the problem. Their advice: assume a 25x load increase within two quarters and design for it now. The full write-up walks through each stopgap.

Why it matters

If your team is adopting agentic coding, the bottleneck may not be the model but the infrastructure around it: CI, test selection, and result storage that were sized for human-paced commit rates. Anthropic's numbers give you a concrete figure to plan against rather than discovering the limit in production.