Image: OpenAI
OpenAI has built a new system called Pando to keep up with the flood of code its engineers and AI agents now produce. The company explained the project in a blog post on November 2 and shared it on X through its OpenAI Developers account.
We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git.
— OpenAI Developers (@OpenAIDevs) November 3, 2026
Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage.
To understand Pando, it helps to know what Git is. Git is the tool most software teams use to track changes to their code. Every saved change is recorded, so developers can see who changed what and go back if something breaks.
OpenAI keeps most of its code in one giant shared collection, known as a monorepo. Sending new code up to that collection is called a push. Downloading the latest code is called a fetch. Pando is the backend, meaning the server side that stores the code and answers those requests.
The problem, according to OpenAI, was Codex. The company's coding agent lets each engineer run many tasks at once, which means far more pushes and fetches than before. OpenAI says that from around the end of last year its old setup struggled, with slow downloads, failed merges and reliability at times falling below 90%.
OpenAI chose to keep Git rather than replace it. Engineers and agents still use the same Git commands as before, and only the system behind them changed.
Pando speeds things up in two main ways. First, it handles each branch on its own. A branch is a separate line of work, such as a new feature someone is testing. Older systems made every update wait in a single line, so one person's push could hold up another's. With Pando, work on different branches can be saved at the same time.
OpenAI admits this has a small cost. Someone looking at several branches at once may see them at slightly different moments in time. The company says that in practice people and agents care about the order of changes within one branch, not across all of them.
Second, Pando reuses downloads. Git bundles code changes into compressed files called packfiles. A normal Git server builds a fresh packfile for each person who asks. Pando instead saves packfiles once and hands the same files to many users, served from fast caches in each region.
Image: OpenAI
OpenAI says the results are large. The typical push now takes about 80% less time, which the blog also describes as nearly five times faster. Pando delivers close to a petabyte of code data a day, which is about one million gigabytes, and has served more than 57 terabytes in a single hour.
The company says Pando was built with Codex in about 10 weeks. One engineer designed the core, teams of AI agents wrote much of the code and a handful of engineers deployed it. It is written in the Rust programming language and began handling real traffic in April. The New Stack reported the same timeline and figures.
OpenAI moved over without downtime by first running Pando alongside its old system and shifting work across in stages. Developers could even ask Codex to move their own setups to the new system.
Building Pando also turned up bugs in Git itself. OpenAI says it fixed problems such as cleaning up after an interrupted download and sent several of those fixes to the official Git project. It tracks that work on a public patch site and in its own copy of Git on GitHub.
Pando is only used inside OpenAI at the time of this writing. The company says it is exploring ways to share the technology with its users, but it has not given a timeline or said what form that would take.
Comments