|
Getting your Trinity Audio player ready...
|
Anthropic says it has moved its entire internal codebase off Git, the version control tool used by almost the whole software industry, and onto a system it built itself. In an engineering post published on October 13, the company described how it designed Scion, a new source control system, and fully migrated its monorepo, the single large repository that holds its code, in five months. The reason, according to Anthropic, is that the volume of code changes grew so quickly that source control became the biggest bottleneck in its development process.
The numbers are the heart of the story. Anthropic said in May 2026 it was merging around 2,500 code changes a day into its main repository. By September, that had passed 30,000 a day, a more than twelvefold jump in four months. The company attributes the surge to the way it now builds software: Claude, its AI model, writes much of its code, and the number of automated agents and jobs touching the repository grew with it. In a post on X, Anthropic said it had become bottlenecked by developer infrastructure built for human scale, not agent scale.
To understand why that breaks Git, consider how Git is designed. It stores a full local copy of a repository’s history on each developer’s machine, and it handles many operations on the client. That works well for teams of people who each work on a few changes at a time. It struggles when thousands of automated clients are constantly fetching, testing and merging against one enormous repository. Anthropic said it first coped by running several hundred Git mirrors, which bought time but grew harder to operate as the merge rate rose, according to a summary of the post by Developers Digest.
Scion takes the opposite approach. Clients fetch only the files they need, while the server handles history, blame, merges and rebases, and nothing requires a fully materialised copy of the repository anywhere, not even centrally, as OfficeChai described it. Internal jobs can query source control through programming interfaces without keeping local repositories, and each tier of the system keeps a cache hit rate above 99 percent without long-lived state, which makes it easy to scale, Anthropic said. The design assumes clients are online, which is true of agents and automated jobs by definition. Anthropic noted that its unusually heavy use of Claude to write software makes its workload different from that of most organisations.
The company reported substantial efficiency gains. Anthropic said Scion’s infrastructure footprint is now 200 times smaller than its old Git mirroring setup in terms of processor and memory use, and that continuous integration and research jobs check out code in seconds instead of minutes. Continuous integration alone freed up 25 percent of processor capacity across the company’s fleet. These are Anthropic’s figures, and the reports reviewed contain no independent measurements.
The team size is the other striking detail. According to the post, a team of seven built both Scion and Antique, a code review tool, plus half a dozen related systems, including code search, a source control permissions layer and new ways of managing and landing large amounts of reviewed code, and fully migrated to them in five months. Anthropic said that in the past, developing and deploying a system like this would have taken years and many more engineers. Digital Today reported that the first version of the pipeline that moved Git data to Scion was mostly written by people, but after the company’s Claude Mythos Preview model rewrote it, memory use fell to one twentieth and speed rose fourfold. Engineers kept responsibility for design and trade-off decisions, and were also responsible for stopping Claude when it broke key constraints, the company said.
The post is also candid about limits. Anthropic said Scion is by no means a suitable replacement for Git for general use, that it only makes sense for enormous monorepos with swarms of agents, and that its internal components are highly bespoke and would be hard for other organisations to adopt. The company said it worries that a public version designed to please everyone would not be very useful, so it does not plan to release one. Instead, it said it expects this class of scaling problem to become common within six to twelve months as more companies lean on agentic coding tools, and that it will work with GitHub, GitLab and the open source community so ideas from Scion, such as server-side remote procedure calls and smarter negotiation protocols, can be folded into Git itself. Readers can find the full account on Anthropic’s engineering blog.
What does this mean for everyone else? For most developers, very little changes tomorrow. Git remains the dominant version control system, and the project’s own documentation and downloads are at Git. The relevance is as an early signal. If AI coding agents multiply the number of changes a codebase receives, the tools around the code, such as version control, code review and continuous integration, can become the limiting factor long before the AI models themselves. A line in the commentary around the news put it neatly: for teams evaluating coding agents, the time spent retrieving, testing and integrating changes deserves attention alongside model speed, because those steps determine how much generated code can actually reach production.
It is worth keeping some skepticism in view. This is a company describing its own engineering, and a workload that is unusual by design: a company whose core product is an AI model that writes code, running at a scale few organisations will match soon. The figures about merges, footprint and speed come from Anthropic’s post, and the reports reviewed do not include outside audits. The shift to a centralised, always-online design also carries trade-offs that Git’s distributed model avoids, such as dependence on the server and reduced ability to work offline, though for Anthropic’s agents those matter less. Whether other organisations hit comparable limits in six to twelve months, as Anthropic predicts, is a forecast, not a fact.
There is a broader point about speed. Seven people built the replacement in five months while the merge rate was rising more than tenfold, which the company presents as evidence that small teams using AI can now do work that once took years. That claim cuts both ways. It suggests productivity gains from AI-assisted engineering, and it also raises questions about review, reliability and security when a codebase changes tens of thousands of times a day. Anthropic’s answer, in its own telling, is a purpose-built code review surface and stronger permissions, but how well those controls hold up at that scale will only be clear over time.
Readers who follow global technology and business stories can find more coverage at BusinessTech Nigeria. For now, the practical takeaway is modest and specific: when software writes software at industrial speed, the plumbing of version control and review stops being a background detail and becomes the story.
