OpenAI built Pando, its own Git backend, after Codex agents flooded its monorepo with commits

OpenAI has described Pando, a Git-compatible backend it built to host its main code repository, in an engineering post titled "Scaling Git at OpenAI". The reason it gives is volume: as coding agents took on more of the work, the number of commits grew faster than its previous Git setup could absorb.

Why OpenAI replaced its Git server

According to AI Weekly's summary of the post, commits landing on the main branch of OpenAI's monorepo rose 1,378% between the first quarter of 2024 and the third quarter of 2026. OpenAI's own projection for the fourth quarter of 2026 is 2,442%, the figure some outlets used in their headlines; since that quarter has only just begun, it is an estimate rather than a measurement. OpenAI attributes the growth to Codex, its coding agent. Under that load, AI Weekly reports, merge reliability on the old system sometimes fell below 90%.

Interesting Engineering reports that Pando has been in production since April 2026. The project went from kickoff to production in about 10 weeks: one engineer designed the core architecture, teams of Codex agents wrote the implementation, and a small group of engineers handled deployment. AI Weekly names Aaron Friel, Taylor Blau and Ted Nyman as the OpenAI staff behind the post.

How it works and what changed

Pando is written in Rust and keeps the standard Git protocol and command-line client, so developers do not change their habits. Rather than ordering every update across the whole repository, it serializes pushes only per branch, letting unrelated branches update in parallel. The trade-off, Interesting Engineering notes, is that a reader looking at several branches at once may see them at slightly different points in time. Incoming packfiles go to cloud object storage, a metadata service tracks branch versions, and previously built packfiles are reused instead of being regenerated for each request.

OpenAI reports that median push latency fell by nearly 80% and merge reliability reached 99.99%. The system serves close to a petabyte of Git data a day, with peaks above 57 TB in a single hour, and more than 99% of bytes come from a local disk cache.

Limits

Pando is an internal tool. AI Weekly notes that OpenAI has published no source code, API or deployment guide, and that the post does not say which Git host Pando replaced or whether the latency gains apply to clone and fetch as well as pushes. All the numbers come from OpenAI. This article could not open the OpenAI post directly and relies on the two reports cited.