1300 166 138 Tell us what you need
News / Anthropic

Anthropic built its own git because Claude codes too fast

By Monika Tantau • Published 1 month ago
A software developer at a desk with six monitors full of code, the kind of work Claude now does at scale inside Anthropic

Key takeaways

• Anthropic merged about 2,500 code changes a day in May and over 30,000 in September
• Git could not keep up, so Anthropic built its own system, Scion, in five months
• A team of seven built it, and its engineers rate Claude as a better coder than themselves
• Anthropic expects other companies to hit the same limits within six to 12 months

Most Read

Anthropic’s engineers now let Claude write so much code that the tools built to manage it broke. The company has replaced git, the system almost every software team uses to track changes, with one it built itself, according to its engineering blog.

The new system is called Scion. Anthropic says writing code is no longer what slows it down. The bottleneck is everything around the code that was designed for humans working at human speed.

From 2,500 to 30,000 changes a day

In May 2026 Anthropic merged about 2,500 code changes a day into its main code repository. By September it was merging more than 30,000, a twelvefold jump in four months. The company does not expect the growth to slow, because every bottleneck it removes leads straight to more code.

The repository itself has become enormous. The history of its main branch alone is almost 100 GB, and its testing systems check out the code around 25 million times a day. Fetched the usual way, that would mean more than 600 petabytes of git traffic every day.

What Scion does differently

Git was designed so that every developer has a full copy of the project and can work offline. Anthropic’s argument is that most of its users are now AI agents and automated jobs that are always online, so that design works against it. Before Scion, it was running several hundred internal git mirrors just to cope with the load.

Scion keeps the project on a central server and lets each agent fetch only the files it needs. Anthropic says the new setup uses 200 times less computing power and memory than the old git mirrors, and that its testing jobs now get their code in seconds rather than minutes.

The GitHub logo on a phone in a jeans pocket, a reminder of git, the system Anthropic replaced with its own tool Scion

Seven engineers, five months

A team of seven built Scion, a new code review tool called Antique and several related systems, and moved the whole company across in five months. Anthropic says a project like this used to take years and many more engineers.

All of the engineers involved now rate Claude as better at writing code than they are. Their job shifted to design decisions and to telling Claude no when it broke the rules of the system or over engineered a fix. They still review every change, but they check the intent of each one rather than reading every line.

Why Australian businesses should care

Few Australian companies produce code at Anthropic’s volume, and Anthropic says as much. But the pattern is already visible in ordinary teams. A Harvard study found that AI coding agents made code reviews take 49% longer, because the extra code piles up at the human checking stage.

Anthropic expects this class of problem to reach other companies within six to 12 months. For a business rolling out tools like Claude Code, the practical lesson is to plan review, testing and deployment for the extra volume before it arrives, rather than measuring success by how much code gets written.

Not a git replacement for everyone

Anthropic is clear that Scion is not meant for general use and will not be released as a product. It only makes sense for giant code repositories with swarms of agents, and it would force a company to replace its code review, search and testing tools as well. Instead, Anthropic is sharing its design so that GitHub, GitLab and the open source community can bring the best ideas to git.

The bigger signal is where the work goes when AI writes the code. Planning for that shift is where AI consulting with software teams now tends to start.

Scroll to Top