30,000 merges a day: what Anthropic's Scion says about git at agent scale
Victoria Shinder3 min readAnthropic's engineering post of 13 October 2026 describes something most teams will never build: an in-house version control system, Scion, to which the company migrated its monorepo from git in five months. The reason to read it anyway is the problem statement, because it names exactly where tools designed for human-paced development start to fail when most commits come from agents.
The numbers behind the decision. Anthropic reports about 2,500 merged changes a day in May 2026 and more than 30,000 a day by September. The history of the main branch alone is close to 100 GB, a compressed working copy over 25 GB, and CI performs on the order of 25 million checkouts a day — which, fetched naively, the post puts at over 600 PB of git traffic daily. Before Scion the company used every standard workaround — shallow and sparse clones, promisor remotes, snapshot seeds in blob storage — and at the peak ran several hundred internal git mirrors.
The observation that changes the design. The post's central argument is that git's strengths assume offline, self-sufficient clients, while nearly all of Anthropic's clients are agents and jobs that are online by definition. From that follow the principles: the server answers history and blame questions and creates commits, merges and rebases; the client keeps only a bounded cache; every repository query is also an API, so many jobs need no local copy at all; commits are server-side "changesets" whose cost scales with the size of the change rather than the repository; refs are discouraged because shared global state is a scaling problem.
What it bought, by Anthropic's measurement. The infrastructure footprint is reported as 200 times smaller in CPU and memory than the previous mirror fleet, checkouts went from minutes to seconds, CI alone gained 25% CPU capacity, and each cache tier holds a hit rate above 99%. A team of seven people built Scion, a code review tool and related systems in that period.
The part that applies elsewhere. Two remarks deserve more attention than the architecture. First, the team says its engineers' main job was often to tell the model "no" — agents violated system constraints to satisfy a pattern, and adversarial reviews pushed towards defensive over-engineering. Second, Anthropic admits it could not keep the usual global lock on the merge queue and moved to post-merge and pre-deploy testing instead, accepting that concurrent changes conflict. Any organisation increasing its share of agent-written code will meet the same question about what CI means when changes merge faster than tests finish.
What the post does not claim. Anthropic says plainly that Scion is not a general replacement for git, that it diverges too far to be adopted by others, and that its own token spending is unusual. Its proposal is narrower: git servers need not mirror git clients, forges could expose server-side RPCs of the kind Scion uses, and the existing sparse and shallow features could be made to compose. The company expects this class of problems to reach more organisations within six to twelve months.
Source: Anthropic Engineering, "Redesigning source control at Anthropic" — https://www.anthropic.com/engineering/redesigning-source-control