On 13 October 2026, Anthropic’s engineering team explained why it stopped using git for its main repository: AI agents now write most of its code and infrastructure designed for people became the bottleneck. It is a clear sign of where software development with AI agents is heading, and it holds useful lessons for any company that wants to automate processes with agents.
What happened: from 2,500 to more than 30,000 changes a day
- In May 2026 Anthropic merged about 2,500 code changes a day; by September, more than 30,000, almost all made by agents.
- Its main repository history is close to 100 GB, and its continuous integration (CI) alone checked out the code about 25 million times a day.
- To keep up, it ran hundreds of internal git mirrors that were increasingly hard to maintain.
Git was designed for people working offline, each with a full copy of the code. Anthropic’s "users" today are agents and jobs that are always online, so those advantages turned into a cost.
What they built
A seven-person team created Scion, its own version control system, and migrated the repository in five months while change volume grew tenfold. Its main ideas:
- Always-online clients: nobody downloads the full repository; each agent requests only the files it needs.
- The server does the heavy lifting: history, commits and merges happen on the server, not on every machine.
- An API for everything: any question about the code is answered through an API, without a local copy.
- Three cache tiers, each with a hit rate above 99%.
According to Anthropic, the new infrastructure uses about 200 times less CPU and memory than the previous one, and CI checkouts went from minutes to seconds.
The role of agents and of people
Claude rewrote one of the team’s prototypes using 20 times less memory and running 4 times faster, found optimizations that would have taken an expert days or weeks, and several instances monitor system components through their telemetry. But people remained key: they had to reject designs that broke core constraints and curb unnecessary defensive code. Engineers now review the intent of each change rather than every line.
Lessons for your company
- Infrastructure built for people breaks at agent scale. A process that works with 10 requests a day can collapse with 1,000. Design for the volume the agent will have, not today’s.
- Give the agent APIs, not screens. An agent that queries your ERP through an API is faster, cheaper and easier to audit than one that mimics clicks.
- People move to setting rules and reviewing decisions. The value lies in clear constraints and reviewing what matters, not redoing every task.
- Controls by area. Anthropic uses per-directory permissions and approval rules so reviewers focus on consequential changes. The same applies to an agent that touches sales, finance or personal data (see AI agent security).
Do you need a system like Scion?
Almost certainly not: the authors themselves say it is not a general git replacement and is built for huge repositories and large groups of agents. What does apply to any company are the principles: data available through APIs, narrow permissions and constant measurement. Anthropic expects this kind of problem to become common within 6 to 12 months.
Source: Redesigning source control at Anthropic (Anthropic Engineering, 13 October 2026).
To bring these principles to your processes, read our AI agent development guide or see our custom AI software service.