When agents write the code, merging is the bottleneck

Anthropic says source control became its biggest bottleneck once Claude sped up coding.

In a post on 13 October, the engineering team gives the numbers. Merges went from about 2,500 a day in May to over 30,000 a day by September. The monorepo history is close to 100 GB, and CI runs roughly 25 million checkouts a day. Git assumes a human who clones a full copy, works for hours and pushes once. Agents don't work like that.

So they built Scion, an internal system where the server does the heavy lifting. Clients stay online and fetch only what they need. Blame and history run on the server. Commits are cheap and don't take locks. Anthropic reports a footprint 200x smaller than the git mirror setup it replaced, and says seven people finished the migration in five months. It also says Scion isn't meant as a general git replacement.

My takeaway: when agents multiply your output, the pipeline around the code breaks long before the model does.

What would break first in your repo if merges went up 10x? Review, CI or the merge queue?