Anthropic reports 30,000 merges a day and says source control, not coding, became the bottleneck

Anthropic says it averaged around 2,500 merged code changes per day in May 2026 and over 30,000 per day as of September. Its monorepo's main-branch history is almost 100 GB, and it runs about 25 million checkouts a day for CI before counting agent sandboxes and research jobs. At peak it operated several hundred internal git mirrors. A seven-person team built an in-house version control system called Scion, plus a code review tool called Antique, and migrated off git in five months. Scion's server handles diffs, blame, commits and merges, so many jobs need no local checkout. Anthropic reports that the new infrastructure footprint is 200x smaller in CPU and memory, that checkouts are two orders of magnitude faster, and that CI alone gained 25% CPU capacity. These are the company's own figures, and the post shows no independent measurement.

The post is as useful on process as on architecture. Reviewers still check every change request to the system, but now "examine the intent of the change as opposed to reading every line of code." Engineers often had to tell Claude no: the authors describe agents that brute-forced merge lookups by reading the commit log, and an adversarial review that pushed Claude into excess validation. On merging, the authors say a proper merge queue "doesn't make much sense at agent scale." They moved to post-merge and pre-deploy automated testing, accept that concurrent changes can conflict, and say Claude can typically fix those conflicts faster than humans can identify them. They also run per-component Claude monitors they call "superfans."

Why it matters: The claim worth testing is that review and CI, not code generation, become the constraint once agents write at volume, and that the answer is checking intent and testing after merge. Anthropic is an extreme case by its own admission, and it says Scion is no replacement for git for general use. Even so, if your agents already produce more changes than your merge queue and reviewers can absorb, this is a first-party description of one team's workable answer, with the caveat that it comes from the vendor that sells the agents.