Pushing 4.7M Commits an Hour
The winning software teams will throw away orders of magnitude more code than they merge.
I ran a benchmark against the custom git server I’m building,
4.79 million pushes per hour to a single repository across 1,408 branches (p50, 322ms).
27.8 million clones per hour of a repository containing the NextJS starter template (p50, 220ms).
4.79M commits won’t make it to master, but the frontier labs are running that many experiments at scale, and the rest of us will be soon.
Anthropic wrote a working C compiler with a swarm of 2,000 Claude Code sessions coordinating over git
Bun’s Rust rewrite did 695 commits/hr for two weeks
DeepMind’s AlphaCode 2 generated 77 million program candidates for 770 submissions in competitive programming.
Cognition is running Agentic MapReduce over their codebase to scan for security vulnerabilities.
Modal can spin up a million sandboxes in seconds, but what’s the point if those sandboxes have nowhere to read or write?
GitHub won’t be the place that the work gets recorded. Branches and repositories should not be sacred. Every agent should use them, liberally. Every parallel agent gets a branch. Every agent turn gets snapshotted and stored in git, so you can provably correlate what turns corresponded which changes.
That’s exactly what I’m building at Corigin. A new git server implementation and wire protocol built from scratch. Distilled down to the essentials that agents need from first principles. Fast, scalable, and easy for agents to use because it’s just git. We benchmarked with git clone and git push.
You can try it today at corigin.dev. Try it with MapReduce skill I made to run sharded reviews or edits over your codebase. Or use it to run parallel cloud agents.