OpenAI built Pando, a Git backend for agent-scale monorepos

On November 2–3, OpenAI Developers (@OpenAIDevs) and OpenAI engineering published that the company built Pando, a new Git-compatible backend for its monorepo.

What Pando is. Per OpenAI's post by Aaron Friel, Taylor Blau, and Ted Nyman (Nov 2, 2026), Pando is a new backend for Git repositories that keeps the standard Git client and protocol. It now powers all development in OpenAI's monorepo. Engineers and agents keep using Git the same way; the backend changed underneath.

Why they built it. Codex multiplied parallel branches, pushes, fetches, and clones from development environments, research jobs, and CI. Starting around the end of last year, the systems behind the monorepo were struggling — slow fetches and pushes, failed merges, reliability at times below 90%. OpenAI wanted horizontally scalable writes (many concurrent pushes to different branches) and high read throughput without abandoning Git.

What it does in production. OpenAI says Pando cut median push latency by nearly 80% (the lede also frames that as nearly 5x faster — same claim). It now delivers nearly a petabyte of Git data a day, with more than 57 TB of Git packs in a single hour and more than 99% of those bytes from local disk cache. Design highlights they name: causal consistency across refs (not a repository-wide point-in-time snapshot), serializable updates per branch, reused packfiles, packfile-uri, and Express regional caches. Some client improvements were contributed upstream to Git.

How it was built. Kick-off to production in about 10 weeks. Designed by one engineer, implemented by teams of agents (Codex), deployed by a handful of engineers. Began serving production traffic in April of this year (2026), first as a read- and write-through proxy, then as the main path.

Who it's for (today). OpenAI engineers and Codex agents working in that monorepo. For MODEL readers: anyone building agent fleets that are already hitting Git push/fetch limits at monorepo scale — this is the architecture OpenAI chose when reliability dipped under agent load.

The catch. Pando is internal. OpenAI says it is "exploring ways to share the technology" — it is not a public product users can install today. MODEL has not tested Pando. Not investment advice.

MODEL Editors