Anthropic says AI coding has made a new problem for its engineers. Writing code is no longer the slow part. The slow part is the tooling that stores, checks and ships that code. In an engineering blog post published on 13 October 2026, the company says it built its own version control system, called Scion, and moved its main code repository off git in five months [1].
What version control and a monorepo are
Version control is the system that records every change made to a piece of software: who changed what, when, and why. It lets many people (or programs) work on the same code, compare versions, undo mistakes and merge their work together. Git is by far the most widely used version control system, and services such as GitHub and GitLab are built on it.
A monorepo (“monolithic repository”) is a single shared repository that holds most or all of a company’s code, rather than splitting it into many smaller ones. Google and Meta famously work this way. It makes it easier to change many projects at once, but the repository can become enormous.
Why Anthropic says git stopped working for it
Anthropic says the volume of code being merged into its monorepo jumped as its Claude models took on more of the coding. According to the post, the company averaged about 2,500 merged code changes a day in May 2026 and more than 30,000 a day by September [1].
“At Anthropic, coding is no longer the bottleneck,” the post says. The bottleneck now, it says, is “all of the developer infrastructure built for human scale: source control, builds, and CI” (continuous integration, the automated testing that runs on each change) [1].
The company gives some figures on scale. It says the history of its main branch alone is almost 100 GB, a compressed copy of the working files is over 25 GB, and it runs on the order of 25 million code checkouts a day for automated testing. Fetched naively, Anthropic says, that would be more than 600 petabytes of git traffic a day [1].
Git was designed so that every copy of a repository is complete and works offline. Anthropic says that is the right design for most projects, but most of its “clients” are AI agents and automated jobs that are always online. It had used the usual workarounds, including partial copies and, at its peak, several hundred internal git mirrors, but says these became harder to run as the merge rate grew [1].
How Scion is different
Anthropic says it drew on earlier large-scale systems, Google’s Piper and Meta’s Sapling, and built Scion around a few principles [1]:
- Clients are always online. Instead of each machine holding the whole repository, clients fetch only what they need from a server.
- The server does the heavy work. The server answers questions about file history, creates commits and performs merges. The client keeps a small local cache.
- Everything scales sideways. There is no single full copy of the repository anywhere; the data is spread out so more servers can be added.
- Commits are cheap. Creating a commit should cost in proportion to the size of the change, not the size of the whole repository.
- Checkouts are optional. Many automated jobs can ask the server questions directly rather than downloading the code at all.
Alongside Scion, the company says it built a code review tool called Antique, code search and an access-control system for the repository in the same five months [1].
Anthropic reports that the new setup uses 200 times less computing and memory than its old git mirrors, that automated jobs now check out code in seconds rather than minutes, and that its testing system gained 25% CPU capacity. These are the company’s own figures [1].
Claude helped build it
The post says a team of seven people built Scion and Antique, and that Claude wrote much of the code. In one example, Anthropic says an early internal version of Claude Mythos Preview rewrote a prototype ingestion pipeline so it used 20 times less memory and ran four times faster [1]. The engineers’ job, the post says, was often to tell Claude “no” when it broke design rules or over-engineered things.
Not a git replacement for everyone
Anthropic says “Scion is by no means a suitable replacement for git for general use” and that it is not planning a public version. It expects other companies using AI coding tools to hit similar problems “over the next six to 12 months” and says it wants to share ideas with GitHub, GitLab and the open-source community so some of them can be applied to git [1].
What this means
- The bottleneck has moved. When AI writes code far faster than people, the systems built around human speed (storage, review, testing) become the limit.
- The numbers are Anthropic’s. The scale and speed figures come from the company’s own post and have not been independently measured.
- Most teams don’t need this yet. Anthropic itself says Scion only makes sense for truly enormous repositories.
Spotted via @AnthropicAI on X: https://x.com/AnthropicAI/status/2110047081414987957
Sources
- Anthropic Engineering, “Redesigning source control at Anthropic”, 13 October 2026. https://www.anthropic.com/engineering/redesigning-source-control
