Anthropic Fired Git. Claude Wrote Too Much Code.
Anthropic replaced Git with a custom system after AI agents pushed code merges from 2,500 to 30,000 a day. Human infrastructure is breaking.
Tools · Source: TechSpot
What happened
Anthropic just ditched Git for its main code repository. They built a new version control system called Scion. The reason is simple. Claude is writing too much code. In May, Anthropic averaged 2,500 merged code changes a day. By September, that number blasted past 30,000. Machine clients now author more than eighty percent of merged code at the company. Git was never designed to handle this kind of volume.
Git was built in 2005 for humans working offline. It assumes every user needs a full copy of the repository history. Anthropic has a massive codebase. Their main branch history approaches 100 gigabytes. A compressed working copy exceeds 25 gigabytes. Continuous integration alone runs 25 million checkouts daily. Downloading full copies for every checkout would generate massive amounts of useless traffic. Anthropic tried shallow clones and internal mirrors. It was not enough.
Seven engineers built Scion in five months. Claude actually helped write the code. An early version of Claude rewrote a data pipeline four times faster. Scion moves heavy operations like merges and file history queries to the server. The client just tracks local edits. It uses one two-hundredth of the CPU power and memory compared to their old Git infrastructure. Checkouts now take seconds instead of minutes. This freed up a quarter of their continuous integration compute capacity.
Key facts
- 30,000 — Merged code changes per day at Anthropic in September
- 25 million — Daily checkouts performed by Anthropic continuous integration
- 1/200th — CPU power and memory used by Scion compared to Git mirrors
- 5 months — Time it took a seven person team to build Scion
- 80% — Minimum share of merged code authored by Claude at Anthropic
Why it matters
The bottleneck in software development is no longer writing code. The bottleneck is the infrastructure built to manage it. Tools designed for human typing speeds break when AI agents submit thousands of pull requests a day. If you build AI coding agents, you have to rethink how they interact with version control. Agents are always online. They do not need offline Git clones. They need fast server side access.
The economic value of source control is shifting from the client to the server. Anthropic proved that a small team can rebuild core infrastructure fast when AI helps. Most companies will not abandon Git completely. But platforms like GitHub and GitLab will have to rebuild their backends to support server side operations. If they fail, agent native platforms will replace them. The moat is no longer just hosting Git. The moat is handling machine scale merges.
For builders
Build server side APIs for agents
AI agents do not need local copies of massive repositories. Build tools that let agents query code and submit changes directly to the server. Companies paying for massive compute overhead will buy solutions that cut continuous integration costs. You win by saving them money on server bills.
Rethink automated testing pipelines
Global repository locks and merge queues fail at agent scale. Hundreds of changes will merge while tests are still running. Engineering teams must invest heavily in post merge automated testing to catch conflicts before deployment. QA platforms that can handle concurrent AI merges will dominate.
Agent native developer tools
The monopoly of legacy code platforms is cracking. Startups have a window to build agent first developer tools. Incumbents who fail to adapt their infrastructure for machine clients will lose enterprise contracts. Founders should target the infrastructure layer that sits between the AI agent and the codebase.
My take
We spent twenty years building tools around human limits. Now AI is shattering those limits. If your product assumes a human is on the other side of the keyboard, you are already obsolete. Build for the machines.
Original reporting: TechSpot. This is my rewrite and opinion.