Anthropic Killed Git: How AI Agents Forced a Source Control Rewrite
Anthropic merged 30,000 code changes a day in September. Git broke under the AI volume, so they built a new version control system.
Tools · Source: Hacker News
What happened
AI writes code faster than human infrastructure can handle. Anthropic hit a massive wall. In May 2026, they averaged 2,500 code changes merged per day. By September, that number exploded to over 30,000 per day. The latent demand for merging code became practically infinite. Every bottleneck they removed just led to more code production. Git simply could not keep up with this machine-driven volume.
To survive, the company built Scion. This is an entirely in-house version control system. They migrated their massive monorepo off Git in just five months. They also shipped a new code review surface called Antique, code search, and a source control IAM system. Claude actually wrote the code that broke the original system. Then Claude helped them quickly build a high-quality, scalable replacement to fix it.
Git assumes developers work offline and need a full copy of the repository. Scion assumes clients are always online agents. The server does the heavy lifting for merges, rebases, and history lookups. Clients just hold a bounded local cache and lay out working files on disk. Commits are cheap, incremental, and lock-free. This architecture drops their infrastructure footprint by 200x compared to their old Git mirror setup.
Key facts
- 30,000 — Code changes merged per day at Anthropic in September 2026
- 100 GB — Size of the history of Anthropic's main branch alone
- 25 million — Code checkouts performed per day for CI at Anthropic
- 600 PB — Estimated daily Git traffic if fetched naively
- 200x — Reduction in infrastructure footprint with Scion compared to Git mirrors
- 25% — CPU capacity saved in CI by switching to Scion
Why it matters
Developer tools built for human typing speeds are about to break everywhere. Anthropic is just the first company to hit this wall because they use Claude to write software more than anyone else. When AI agents write your code, your bottleneck shifts entirely. It moves from human thought to continuous integration, builds, and source control. You will need infrastructure designed for machines. Tools built for human scale will strain and collapse under the weight of autonomous research loops and micro-optimizations.
Decentralized version control might become a legacy concept for massive AI teams. Git won the version control wars because it let humans work offline. Agents do not work offline. They live in data centers and are online by definition. Centralized, API-first version control makes more sense when your primary developers are always-online AI models. Jobs can ask a server for a diff without ever downloading a repository. This fundamentally changes how we think about the software development lifecycle.
For builders
Audit your CI and build pipelines now
Your current CI setup assumes human commit rates. If you deploy AI coding agents, your merge volume will scale tenfold rapidly. Prepare to scale your runner capacity or you will pay massive cloud bills for stalled queues.
Stop fetching full repositories
Anthropic performs 25 million checkouts a day. Fetching full Git history for every CI job is a massive waste of compute. Switch to sparse checkouts or virtualized checkouts to save time and money.
Build API-first developer tools
Scion exposes all source control queries via API. Agents do not need a local repository to check file history or blame. Tool builders who decouple repository data from local disk storage will win the AI agent market.
My take
We are entering the post-Git era. Linus Torvalds built Git for distributed humans. I build for centralized agents. If your AI agents are bottlenecked by human-scale tools, you are losing the race.
Original reporting: Hacker News. This is my rewrite and opinion.