OpenAI's Pando Git backend cuts push latency ~80% for Codex
Codex's agents reportedly drove commits up 1,378%, so OpenAI rebuilt Git's backend in ~10 weeks — a signal your version-control host is the next wall.
Copy markdown
A Git backend that keeps Git
Pando is a drop-in server for OpenAI's monorepo: engineers and agents keep running plain git push and fetch, but push latency falls ~80% and it now moves nearly a petabyte of Git data a day. No client changes, no new workflow — Git capacity now scales with agent usage.
Agents, not people, broke Git
Codex reportedly pushed commit volume up ~1,378%, enough that a vanilla Git server became the bottleneck; OpenAI says a small team shipped Pando in about 10 weeks. The takeaway for heavy-agent repos: your Git host, not the model, is the next thing to fall over.
You can't run Pando — but the patches are public
Pando itself stays internal, but OpenAI maintains a public fork of Git (openai/git) carrying its Codex-oriented client patches, with downloadable release artifacts. If your agents commit constantly, treat Git-server scaling as a first-class problem now, not after it melts.