Agentic coding grew Anthropic's CI 25x. Three fixes failed first.

Anthropic's CI job volume grew 25x in six months after Claude started writing most of their code.

What stands out is the order of their fixes more than the number itself:

Fix	How long it held
Double the CPU cores	70 days
Shard tests by package	29 days
Restart CI daily	< 1 day

Each fix bought less time than the one before, because each one treated runtime as the problem. The real problem was volume: the full suite ran on every PR no matter what it touched.

What finally worked was a deterministic test-impact-analysis service. A listener records results from every run, and a selector picks only the tests a given PR needs. They made it stateless so it scales horizontally.

The question to ask about your own CI

Is your pain runtime-bound or volume-bound? If every hardware or parallelization fix buys you a shorter reprieve than the last, you've got the same signal Anthropic got twice.

(Caveat: these are self-reported numbers from one company about its own product. They show one real case, not a benchmark.)

I wrote up the full breakdown, including what the original post doesn't tell teams without a spare senior engineer for three weeks. Link is in the first comment. 👇