Agent-Written Code Drives a 25-Fold Increase in Anthropic’s CI Workload in Six Months, Forcing Its Test Selection Service to Adopt a Stateless Architecture

Anthropic disclosed its experience scaling its internal continuous integration system on September 14: after agents began writing, reviewing, and submitting code at scale, the number of CI jobs increased 25-fold within six months, while total test volume grew tenfold. The company also said Claude writes about 80% of merged code and that engineering output is eight times the 2021–2025 average. These are internal statistics that have not been independently audited, and lines of code cannot be equated directly with productivity.

The component that actually failed was the test impact analysis service. In the old version, a listener sequentially recorded each CI result, after which a selector determined which tests each pull request should run based on historical failures and package dependencies. Because the historical state for each test remained in a single process, the system could not simply add more writers. When the listener fell 20 minutes behind, tens of thousands of results might not yet have been reflected, causing the selector to keep running known flaky tests or temporarily miss newly added or recently fixed tests. This did not mean untested code was deployed directly to production, but it reduced the timeliness and reliability of the selected test suite.

The team successively scaled up the host, sharded by package, and introduced daily restarts. Those three fixes lasted only 70 days, 29 days, and less than one day, respectively. The final version moved state into an in-memory database: any stateless listener can append results to a journal, while a separate lightweight consumer aggregates them every few seconds into per-test histories, which the selector queries directly. The architecture is more expensive, but it can scale horizontally and supports measuring each stage independently. Anthropic said one engineer completed the work in three weeks.

For teams adopting coding agents, the key lesson is that capacity planning cannot focus only on model tokens. Pull request granularity, overnight submissions, the rate of test creation, and automated reviews can collectively amplify downstream load. The next thing to watch is whether Anthropic publishes absolute workload figures, test-selection recall, end-to-end latency, and cost. Teams implementing a similar architecture should also monitor the conservation of input and output events, queue lag, and use sampled full test-suite runs to verify that the selector is not missing regressions.