Claude writes so much code at Anthropic that git couldn’t keep up, so it built its own

Anthropic now merges more than 30,000 code changes a day, up from about 2,500 in May, and much of that code is written by Claude. Git, the system almost every software team uses to track its code, could not keep up, so Anthropic built its own. Scion, an in-house version control system, has replaced git across the company’s monorepo, the single giant repository that holds its code, after a five-month migration by a team of seven.

The company says its bottleneck is no longer writing code but “all of the developer infrastructure built for human scale: source control, builds, and CI”. In the same five months the team also built a code review tool called Antique, code search and a new permissions system.

Why git ran out of road

Git is designed so that every copy of a repository is complete and works offline. That suits most projects, but the history of Anthropic’s main branch is now almost 100 GB and a compressed working copy is more than 25 GB. The company runs around 25 million code checkouts a day for automated testing (CI) alone, before counting developer machines, agent sandboxes and research jobs. Fetched naively, it says, that would be more than 600 petabytes of git traffic a day.

It tried the standard workarounds, including shallow clones, snapshots and, at its peak, several hundred internal git mirrors. They bought time but grew harder to run as merges climbed.

But almost all of its users were agents and jobs that are always online, so the offline design that made git the standard was working against it.

How Scion works

Scion turns git’s model around. The server does the heavy lifting: it creates commits, runs merges and rebases, and answers questions such as who last changed a line. The client keeps only a small local cache and fetches files when they are needed. Everything the client can ask is also available through an API, so many CI jobs never need a copy of the code at all.

Each commit, which Scion calls a changeset, records only the files that changed, with pointers to their contents in cloud storage, while the metadata sits in Google’s Spanner database. It borrows lessons from Google’s Piper and Meta’s Sapling, built for those companies’ own giant codebases.

Files reach machines through three tiers of cache, one on each machine, one in each region and the main servers behind them, and each tier answers more than 99% of requests on its own.

200 times less hardware

Anthropic says Scion needs 200 times less CPU and memory than its old git mirrors, and that faster checkouts have freed up 25% of CI’s CPU capacity. In its own measurements a CI checkout falls from 31 seconds to 3.5, setting up a fresh workspace for an agent from 7.1 minutes to 5.1 seconds, and total source control time for each agent change from 34 minutes to 3.7. By its estimate that saves about 11,000 agent hours and 280,000 machine hours a day. These are Anthropic’s own figures.

“The role of the engineers was to tell Claude ‘no’”

Claude did much of the building. Anthropic says it gave an early internal version of Claude Mythos Preview its largely handwritten prototype of the pipeline that imports git history into Scion and asked it to rewrite the whole thing. With no further instructions, the new version used 20 times less memory and ran four times faster.

Every engineer on the project now rates Claude as a better coder than themselves, the company says, but their judgement still counted. “Very often, we found that the role of the engineers was to tell Claude ‘no,’” the team writes, describing agents that broke core system rules and adversarial reviews that pushed Claude into over-engineering defences against imagined problems. The engineers still review every change, but check what it is meant to do rather than reading every line.

The team also keeps Claude “superfans” in Slack, each watching one part of the system for anything odd.

Not coming to your company, yet

Anthropic isn’t releasing Scion. It says the approach only makes sense for “truly enormous monorepos” with tens of thousands of engineers or “swarms of agents”, and that switching would mean replacing code review, code search and CI as well. Instead it wants to work with GitHub, GitLab and the open-source community to bring some of Scion’s ideas to git. It expects others to hit the same limits “over the next six to 12 months”.

Why it matters

Companies such as Barclays are giving AI coding tools to thousands of engineers, and Anthropic’s numbers show what comes next: once code is cheap to write, everything around it becomes the bottleneck. Seven people replacing git in five months is also the kind of acceleration that led a Google DeepMind engineer to quit last month.

Sources: Anthropic Engineering: Redesigning source control for agent-scale development.