Anthropic Replaced Git With Scion After Claude Pushed Merges to 30,000 a Day

Anthropic has stopped using Git for its main code repository. In an engineering blog post published on October 13, 2026, the company described how it built an in-house version control system called Scion and fully migrated its monorepo off Git in five months. A team of seven people did the work. In the same period, the team also built a code review tool called Antique, plus code search, access control, and code-landing tooling.

The reason is volume. In May 2026, Anthropic averaged around 2,500 code changes merged per day. By September, that figure was over 30,000 per day. Anthropic’s own chart puts the peak at 30,401 merges on September 9. The company attributes the jump largely to Claude-powered coding agents writing and submitting changes at machine speed. It says it does not expect the growth to slow, because latent demand for merging code is practically infinite, and every bottleneck it removes leads to still more code.

The company’s framing is blunt: coding is no longer the bottleneck. The bottleneck is the developer infrastructure built for human scale — source control, builds, and continuous integration. That describes what happens when code becomes cheap to produce. The constraint moves downstream, to the systems that store, review, test, and merge the output.

The numbers that broke Git

Anthropic’s monorepo is large by any measure. The history of its main branch alone is almost 100 GB. A compressed working copy is over 25 GB. For continuous integration alone, the company runs on the order of 25 million code checkouts per day, before counting developer machines, agent sandboxes, and research jobs. Fetched naively, Anthropic says, that would be over 600 petabytes of Git traffic per day.

It had tried the standard fixes. Shallow clones, snapshots, and a fleet of internal Git mirrors bought time. At its peak, Anthropic operated several hundred internal Git mirrors. As merge volume climbed, those workarounds grew harder to run. The mirrors needed maintenance, and every checkout still had to negotiate with a history most jobs would never read.

The deeper problem was a mismatch of assumptions. Git is designed so every clone is a complete copy that works offline. That design is pleasant on a laptop. It is wasteful when nearly all clients are agents and automated jobs that are online by definition and usually need one directory or one diff, not the full history of the company.

How Scion works

Scion turns Git’s model around. It assumes clients are always online and puts the heavy lifting on the server. The server 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.

Anthropic calls each commit a changeset. 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. Directory trees are derived by the server rather than stored.

Checkouts are sparse by design and can be virtualized, so a job fetches only the files it actually reads. Files reach machines through three tiers of cache — one on each machine, one in each region, and the main servers behind them — and Anthropic says each tier answers more than 99% of requests on its own. The design draws on prior art for huge monorepos, notably Google’s Piper and Meta’s Sapling.

The surrounding tooling changed too. Antique is the new code review surface, replacing Git pull requests with change requests. Anthropic also rebuilt code search, source control access control, and the tooling that lands and merges changes at scale.

What the switch bought

Anthropic’s measured gains are large. It says the Scion infrastructure footprint is 200 times smaller than its previous Git mirroring setup in CPU and memory requirements. CI and research jobs check out code two orders of magnitude faster, in seconds rather than minutes. That freed capacity across the fleet, with CI alone gaining 25% in CPU capacity.

The per-task numbers make the change concrete. In Anthropic’s own measurements, a CI checkout fell from 31 seconds to 3.5 seconds. Setting up a fresh workspace for an agent fell from 7.1 minutes to 5.1 seconds. Total source control time for each agent change fell from 34 minutes to 3.7 minutes. By Anthropic’s estimate, that saves about 11,000 agent hours and 280,000 machine hours a day. These are the company’s own figures, not independent measurements, but the direction is consistent with the design: stop moving full histories to jobs that need a handful of files.

Claude built part of its own replacement

Claude did not only create the bottleneck. It also helped build the replacement. 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 rewritten version used 20 times less memory and ran four times faster.

The human role shifted in the process. Every engineer on the project now rates Claude as a better coder than themselves, the company says. But their judgment still counted. The team writes that very often, the role of the engineers was to tell Claude “no” — describing agents that broke core system rules and adversarial reviews that pushed Claude into over-engineering defenses against imagined problems. Engineers still review every change, but they check what it is meant to do rather than reading every line.

Not coming to your repo

Anthropic is not 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, the company wants to work with GitHub, GitLab, and the open-source community to bring some of Scion’s ideas to Git. It expects other companies to hit the same limits over the next six to twelve months.

That last point is the one to watch. Anthropic’s numbers are extreme, but the shape of the problem is general. Any team running coding agents at scale will produce more changes than its review, test, and merge pipeline was built to absorb. The first symptom may not be slow code generation. It may be slow checkouts, crowded review queues, or CI that cannot keep up.

What this means for builders

For most teams, the practical lesson is not to replace Git. It is to notice where the bottleneck moved. If agents write most of your code, the expensive parts become fetching code, reviewing changes, running tests, and landing merges. Sparse checkouts, API-first access to repository data, aggressive caching, and review tooling that can handle machine-speed volume matter more than they did when humans were the main authors.

Source: Anthropic Engineering blog post, October 13, 2026.