OpenAI's Pando Git backend cuts push latency by 80%

OpenAI published details on 2 November of Pando, a custom Git backend it built for its internal monorepo after Codex-driven commits pushed the volume of pushes and fetches well past what its previous setup could reliably serve, with merge reliability sometimes dipping below 90%.

Pando keeps full compatibility with the standard Git protocol and existing clients, so nothing changes for a developer or an agent running ordinary Git commands. Underneath, it uses causal consistency rather than one global ordering, letting writes to unrelated branches proceed independently, and stores packfiles in object storage with regional caching. OpenAI says the system now serves nearly a petabyte of Git data a day, has cut median push latency by close to 80%, and at peak has handled more than 57TB of pack data in a single hour, over 99% of it served from local disk cache. It took one engineer about ten weeks, with Codex agents doing much of the implementation, and has been in production since April.

Why it matters: this is a concrete data point on what agentic coding does to infrastructure that was never built for it. If coding agents are pushing and rebasing at a rate no human team would generate, your CI, your Git hosting and your review tooling are the next things to feel it, not just your editor. OpenAI isn’t open-sourcing Pando itself, but says it’s contributing fixes back to the Git project.

The caveat: this is OpenAI describing its own infrastructure at a scale most teams will never approach. The lesson is the direction, not a backend to go and copy.