Twelve times the merges in four months
Anthropic has moved its internal monorepo off git and onto a version control system it built itself, called Scion. In an engineering post published on Monday, the company says the repository was merging about 2,500 changes a day in May 2026 and more than 30,000 a day by September — a twelvefold rise in four months, driven by Claude writing most of the code.
The scale numbers are what make the case. The main branch’s history runs to almost 100GB, and a compressed working copy is over 25GB. Continuous integration alone pulls roughly 25 million checkouts a day, before developer environments, agent sandboxes and research jobs are counted. Fetching those naively over git, Anthropic says, would mean more than 600 petabytes of traffic a day.
The workaround had been git mirrors — several hundred of them at peak — and they became the thing that broke.
What Scion actually changes
Scion moves the work to the server. History, blame, commits and merges live there and are reached over RPC; clients fetch only what they need, sparsely and shallowly. Git’s design assumption — a full local copy, offline-first — fits badly in a repository where nearly every client is an always-online agent or automated job.
Anthropic reports the new infrastructure footprint is about 200 times smaller in CPU and memory than the git mirroring it replaced, that CI and research checkouts are roughly two orders of magnitude faster (seconds rather than minutes), that the checkout gains freed about 25% of CI CPU capacity, and that cache hit rates exceed 99% at each tier. These are the company’s own measurements of its own system.
Seven people built Scion, a code review tool called Antique, and several related systems, and completed the migration in five months.
“The role of the engineers was to tell Claude ‘no’”
The post is unusually direct about how it was written. Anthropic says the engineers who worked on Scion agree Claude writes better code than they do, and that their job became review. One line stands out: “Very often, we found that the role of the engineers was to tell Claude ‘no’.” It also notes that “Agents will sometimes violate core system constraints to satisfy a pattern” that should not have been allowed.
One concrete example: an early git-to-Scion ingestion pipeline was hand-written as a prototype, and the version Claude produced used 20 times less memory and ran four times faster.
Not for you
Anthropic is explicit that this is not a general recommendation: “Scion is by no means a suitable replacement for git for general use.” It is not being open-sourced, because it diverges far enough from git to require its own code review, code search and CI to be useful at all. The company frames the post as a case study for very large monorepos — tens of thousands of engineers, or heavy agent-written code — and says it hopes git platforms and the open-source community adopt parts of the design, such as server-side RPCs and sparse, shallow fetching, without abandoning git’s model.
The Register covered the post on Tuesday. The checkable thing from here is whether any git platform ships server-side changes along these lines, because Anthropic’s merge curve is not unique to Anthropic — it is what happens to any codebase where the agents outnumber the engineers.
