Anthropic says its engineers merged about 2,500 code changes a day in May and more than 30,000 a day by September. Git was not losing commits. The surrounding machinery—mirrors, checkouts, ancestry queries and CI—was buckling under a workload designed for agents. Anthropic replaced Git for its monorepo with Scion, a server-side version-control system built in five months.
The bottleneck moved downstream from writing code
The numbers in Anthropic’s engineering account explain why normal fixes stopped being enough. Main-branch history approached 100GB, compressed working copies exceeded 25GB and CI performed roughly 25 million checkouts a day. Naively fetched, that workload would create more than 600 petabytes of daily Git traffic. Anthropic had already built hundreds of internal mirrors.
Scion centralizes operations that Git normally lets every clone perform. A small server fleet answers requests for changesets, diffs, merge bases and blobs. Content-addressed objects never change, so cache layers can serve them without coordinating every read. Clients request only the trees and files they need instead of negotiating a giant packfile.
Anthropic reports cache hit rates above 99 percent, checkouts falling from minutes to seconds, infrastructure using 200 times less CPU and memory than its old Git mirrors, and a 25 percent CPU-capacity gain in CI. These are company measurements from an unusual internal environment, not a benchmark saying ordinary teams should migrate.
Claude created the pressure and helped build the release valve
Seven engineers built Scion and its Antique review surface while merge volume increased tenfold. Anthropic says an early Claude Mythos model rewrote its ingestion pipeline to use 20 times less memory and run four times faster. The team also says engineers still review every change, but increasingly examine intent instead of every line.
That last shift matters more than the new database. AI coding agents can create plausible changes faster than humans can establish whether those changes belong. Anthropic says engineers often had to tell Claude “no” when it violated a system constraint, chose a brute-force path or over-engineered a response to hypothetical criticism. More code made architectural judgment scarcer.
Do not replace Git because Anthropic did
Git’s full local history and offline operation are strengths for almost every repository. Scion trades those properties for centralized scale inside one enormous monorepo. Unless your CI runs millions of checkouts or your mirror fleet is a line item, the migration cost would be absurd.
The transferable lesson is to measure the whole delivery system. If agent output rises but review queues, tests and deployments do not, more generated code will not produce more software. A recent study summarized by AI Weekly found exactly that pattern across hundreds of companies: more code and pull requests, with finished issues largely unchanged and review time rising.
Start with merge latency, failed CI retries, review time and duplicate work. Then fix the limit you actually have. Anthropic needed new source control. Most teams need smaller change sets, clearer ownership and enough reviewers to decide what should ship.
Cover art: AI-generated editorial composition by TMRW.
