OpenAI details Pando Git backend after Codex commit surge

Commits landing on master at OpenAI grew 1,378% between the first quarter of 2024 and the third quarter of 2026, and the company's own estimate for the fourth quarter is 2,442%. Codex, the agent writing most of that code, pushed the old setup past breaking, with reliability at times falling below 90%.

In a post on openai.com, members of technical staff Aaron Friel, Taylor Blau and Ted Nyman describe the fix. They built Pando, a new backend that preserves standard Git clients and workflows while rewriting what happens on the server. "It was clear that we needed to revamp our source control system to scale to the level of concurrent work that Codex makes possible," they write.

Two moves do the work. Pando trades repository-wide consistency for causal consistency, letting pushes to different refs commit in parallel while pushes to the same ref stay serialized. "A client reading several branches cannot assume those reads describe a single point in time across the repository," the authors note. And it reuses cached packfiles across clients instead of generating one per request, so "the server can serve the same bytes to many clients with little computation."

The production numbers OpenAI puts on it: median push latency down nearly 80%, merge reliability at 99.99%, close to a petabyte of Git data served per day, and more than 57 terabytes in a single hour with over 99% of those bytes from local disk cache. The company says the Rust backend went from kickoff to production in 10 weeks, with one engineer designing the core, teams of agents implementing it, and a handful of engineers deploying it.

Pando remains an internal system; OpenAI has published no source code, no API, and no deployment guide. The claims are self-reported, and the post does not say which prior Git host Pando replaced. Among the hundreds of AI infrastructure stories we have tracked in the last ninety days, almost none have concerned source control plumbing.