Anthropic says agents pushed it to over 30,000 merged code changes a day, so it built its own source control system and left git in five months
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months:
Context
Anthropic's internal numbers show how fast AI agents are multiplying software output, and that tools built for human scale are now breaking under it.
This is Anthropic's own engineering blog, so it is first hand. The company says its token spend for internal use far exceeds what most businesses would spend, so its workload is unusual. It says the system, called Scion, is not a general replacement for git.
Deep dive
On October 13, Anthropic posted that in May it merged around 2,500 code changes per day and in September more than 30,000 per day, leaving it bottlenecked by developer infrastructure built for human rather than agent scale.
The engineering blog says Anthropic built Scion, an in house version control system, and fully migrated its monorepo off git in five months, along with a code review tool called Antique, code search and access controls. It says the main branch history is almost 100 GB, that it performs on the order of 25 million checkouts a day for CI, and that at its peak it ran several hundred internal git mirrors.
Scion assumes clients are always online, so the server handles history, blame, commits and merges, and clients fetch only what they need. Anthropic says its Scion infrastructure is 200 times smaller in CPU and memory than the old git mirrors, checkouts are two orders of magnitude faster, and CI gained 25% CPU capacity.
The blog says a team of seven built Scion and Antique, and that engineers now rate Claude as better at writing code than themselves, while their role is often to tell Claude no. It says an early Claude Mythos Preview rewrote one pipeline to use 20 times less memory and run four times faster.
What to expect next
Watch whether GitHub, GitLab and the git community adopt ideas Anthropic recommends, such as server side APIs and better sparse checkouts, which it says it wants to partner on.
Also watch whether other heavy users of coding agents report similar bottlenecks in source control, builds and CI over the next six to 12 months, as Anthropic predicts. Anthropic's merge numbers will also be a useful signal of how quickly agent written code is growing inside the largest AI labs.