Anthropic Scion Replaces Git as Merges Jump From 2,500 to 30,000 a Day

Anthropic Scion is the in-house version control system that Anthropic built to replace git for its monorepo, and the reason is a number that sounds made up. In May 2026 the company averaged around 2,500 merged code changes per day. By September it was merging more than 30,000 per day, mostly written by Claude. Anthropic described the five-month migration in an engineering post published October 13.

What actually happened

Anthropic's post opens with a blunt claim: coding is no longer the bottleneck, developer infrastructure built for human scale is. Source control, builds and CI all started to strain once Claude was writing and landing code faster than any human team. The monorepo's main branch history alone is almost 100 GB, and a compressed working copy is over 25 GB.

The checkout load is the wild part. Anthropic says it performs on the order of 25 million code checkouts per day for CI alone, before counting developer environments, agent sandboxes and research jobs. Fetched naively with git, that would be over 600 PB of traffic per day. At its peak the company ran several hundred internal git mirrors just to cope, and those mirrors kept falling behind as the merge rate climbed.

So the team asked a simple question: if nearly every client is an agent or a job that is online by definition, why use a system designed for offline clones? The answer became Anthropic Scion, plus a code review tool called Antique, code search, a source control access system and new merge tooling, all built and shipped in the same five months.

How Scion differs from git

Git gives every clone the full repository so it can work offline. Anthropic Scion flips that. The server answers questions about history and blame, creates commits and performs merges. The client lays out files, watches for changes and fetches only the trees and blobs it actually needs. Sparse and virtualized checkouts are the default, not an add-on.

The unit of work is a "changeset" in a write-ahead log, storing the files that changed plus content-addressed pointers to compressed contents in blob storage. Because everything is content-addressed, any cache can serve any object, and Anthropic says each cache tier holds a hit rate above 99%. Refs, like branches and tags, are discouraged except for main and release branches, because global state every server must agree on becomes a scaling trap.

Very often, we found that the role of the engineers was to tell Claude "no."

That line from the post is the most honest part of the story. Anthropic says agents sometimes violated core system constraints, and skilled engineers had to steer Claude away from over-engineering in response to hypothetical concerns raised by adversarial review.

The Claude angle

Anthropic Scion is also a showcase for the models. The post says an early internal snapshot of Claude Mythos Preview rewrote the handwritten git-to-Scion ingestion pipeline with no further instructions, producing a version that used 20x less memory and ran 4x faster. A team of seven built Scion and Antique and migrated the company in five months while merge volume grew roughly tenfold.

The team also runs Claude "superfans" in Slack channels, each one obsessively watching the telemetry of a single component. Anthropic says CI alone gained a 25% CPU capacity win from faster checkouts. All of these figures are self-reported, and The Register's write-up repeats the same numbers rather than independently measuring them.

If this sounds familiar, it rhymes with the agent swarms in our Claude dynamic workflows coverage: once one agent can spawn many, the bottleneck moves to whatever infrastructure has to absorb the output. It also sits awkwardly next to Claude unintended actions, a reminder that more autonomous code still needs humans who know when to say no.

What to do as a reader

If you run engineering at a normal company, do not rip out git. Anthropic says so itself: Scion diverges heavily from git, needs its own review, search and CI, and would be hard to package for other organizations. The company does expect this class of problem to become common over the next six to 12 months, and says it wants to work with GitHub, GitLab and the open-source community so git servers can adopt some of the same ideas.

The practical takeaway is to watch your own merge rate and CI checkout cost as agent usage grows. If those curves start bending the way Anthropic's did, server-side APIs, sparse checkouts and post-merge testing are the levers the post highlights. For investors following the company itself, our Anthropic Q3 financials note covers the IPO chatter. Anthropic Scion is an internal tool story, not a product launch, and nothing here is investment advice.

Primary source: Anthropic's engineering post, with a summary from The Register.