Anthropic replaced git with Scion after AI pushed it to 30,000 code changes a day

Anthropic says AI coding broke git. In May its engineers merged about 2,500 code changes a day. By September it was merging more than 30,000 a day, and the company credits Claude's coding for the jump. Git, the standard version control tool that GitHub is built on, could not keep up at that volume.

So seven people built a replacement called Scion in five months, along with an in-house code review tool. The load numbers explain why. Anthropic's automated tests grab a copy of its code roughly 25 million times a day, and at the peak the company was running several hundred copies of its git servers just to keep up.

Scion's core idea is to stop loading the whole repository. Systems fetch only the parts a task needs, which Anthropic says cuts compute use sharply and turns code preparation for automated checks and research from minutes into seconds. The company's framing is blunt: at this scale, writing code is no longer the bottleneck. Storing, searching, building and reviewing it is.

Our take

This is the clearest evidence yet that AI coding moves the bottleneck from writing code to everything around it. Version control, CI, builds and review were sized for human output.