Anthropic has replaced Git in its main code repository with Scion, a system designed around a startling change in scale: the company says merged code changes grew from about 2,500 a day in May to more than 30,000 by September. The system shifts much of source control onto servers, cutting the data agents and automated jobs need to fetch.
Anthropic Watch analysis
What happened
TechSpot reports that Scion handles commits, merges, rebases and file-history queries on the server. Clients keep local edits and a limited cache; jobs can fetch selected directories or retrieve files as they are read, rather than downloading a full repository copy.
Anthropic says its main-branch history approaches 100GB and that continuous integration performs about 25 million checkouts a day. The company says Scion uses one two-hundredth of the CPU power and memory of the Git mirror infrastructure it replaced, makes checkouts take seconds instead of minutes, and freed 25 per cent of its CI CPU capacity. Those figures are Anthropic’s, as reported by TechSpot. Read TechSpot’s report.
Why it matters
This is a practical consequence of AI-assisted software development, not just a larger tally of code. When agents and automated jobs generate changes at this volume, the tools around a repository can become a bottleneck too. Scion is Anthropic’s answer to the repeated storage, synchronisation and checkout work involved in a very large, busy codebase.
The design suits Anthropic’s mostly online agents and jobs, but it trades some local independence for centralised server operations. Anthropic says the system would be difficult for other organisations to adopt because it relies on custom review, search and CI tools. Git remains suitable for most repositories, the company says.
Our read
The interesting signal is that AI-assisted coding can change the infrastructure needed to manage software, not only how quickly someone writes a function. Scion is a substantial response to Anthropic’s own workload, not a ready-made instruction for every engineering team to abandon Git. The useful lesson to borrow may be narrower: if agents repeatedly fetch far more code than they need, server-side operations and selective checkouts deserve a look.
What to watch
- Whether Anthropic shares more about Scion’s reliability and how it handles busy periods.
- Whether existing version-control platforms adopt server-side features that reduce repeated code downloads.
- Whether other teams report similar infrastructure pressure as agent-generated changes increase.
Discussion spark: If AI agents are generating code at this scale, should software teams redesign their source-control infrastructure around agents, or keep improving the tools built for human developers?
Sources and evidence
- Source update (15 October 2026, 13:43 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.