OpenAI Pando is a new Git-compatible backend that OpenAI built because its Codex coding agents were pushing code faster than regular Git could handle. It cut push latency by nearly 5x and now serves close to a petabyte of Git data every day. You can’t download it, though, because it isn’t open source.
Here’s what Pando does, why OpenAI needed it, and what it means if your team is starting to lean on AI coding agents.
These are the facts worth knowing before we go deeper:
- Pando is OpenAI’s own Git backend, written in Rust and in production since April.
- Push latency dropped by nearly 5x, according to The New Stack.
- It serves close to a petabyte of Git data per day.
- One engineer designed the core, and Codex agents wrote much of the code in about 10 weeks.
- It is not open source, and OpenAI says it is still exploring ways to share the technology.
The short version: when agents write code, the bottleneck moves from the developer to the version control server.
What OpenAI Pando actually is
Pando sits where GitHub-style servers usually sit. Your Git client talks to it just like any other Git server, so nothing changes on the developer side.
What changed is the server design underneath. OpenAI says the real limit was in how conventional Git servers work, not in the Git client or the protocol, so it kept compatibility and rebuilt the part in the middle.
That choice matters. Teams don’t have to learn a new tool, and existing scripts and CI jobs keep working.
Why Git started breaking under Codex
Agents don’t work like people. A developer might push a few times a day, while a fleet of agents can push constantly, in parallel, across many branches.
The New Stack reports that OpenAI’s monorepo commits are projected to rise more than 2,400% by the end of 2026, compared with the first quarter of 2024. At the same time, reliability of the old Git backend had fallen below 90%.
In practice, that meant slow fetches and failed merges. Every failed push is an agent sitting idle, which wastes compute you’re paying for.
We saw the same trend on the GitHub side, where AI agents now start over 40% of pull requests. More agent traffic means more pressure on the Git layer.
How Pando speeds up pushes and fetches
The core idea is simple: stop making unrelated work wait in the same line. This table shows the main differences.
| What matters | Conventional Git server | OpenAI Pando |
|---|---|---|
| Ref updates | Coordinated across the whole repository, so one branch can wait on another | Handled per branch, so pushes to different branches don’t block each other |
| Fetches | Packs are often generated on demand for each fetch | Serves stored packs instead of building them each time |
| Caching | Varies by setup | Regional caches (called Express) hit above 99% |
| Compatibility | Standard Git | Still Git-compatible, so clients and protocol stay the same |
Why this matters: a push to one branch no longer stalls because someone else is busy on another. For agent-heavy teams, that removes a lot of dead time.
The reported numbers back it up. The median pack in the monorepo is under 20 KiB, so most of the traffic is many small, frequent changes. Serving stored packs fits that pattern well.
Can you use OpenAI Pando yourself?
Not today. OpenAI hasn’t released Pando as open source or as a product, and it only says it is exploring ways to share the underlying technology.
It is sharing some of the groundwork, though. Several Git fixes, such as improvements to packfile-uri handling, are being sent upstream, and some are already merged. Others sit in OpenAI’s public Git repository with a public patch tracker.
So even if you never touch Pando, regular Git may get a little faster because of this work. You can read OpenAI’s own write-up, Scaling Git at OpenAI, for the engineering detail.
What this means for teams using AI coding agents
If you run a handful of agents on a small repo, you won’t hit these limits. The problem shows up at scale, with a big monorepo and many agents working at once.
Still, it’s a useful warning. Agent-heavy workflows can strain tools that were built for human pace, from Git servers to CI queues to code review. Even the limits on the AI side are changing, as in our look at ChatGPT and Codex sharing one usage limit.
Here’s what to watch in your own setup:
- Push and fetch times during busy hours, not just on average.
- Failed merges or rejected pushes that rise as agents join the team.
- CI and review queues, which often choke before Git does.
Why this matters: spotting the bottleneck early is cheaper than adding more agents on top of a slow pipeline.
Frequently Asked Questions
What is OpenAI Pando?
It’s OpenAI’s in-house, Git-compatible backend, built to handle the heavy parallel traffic from Codex agents. It entered production in April.
Is OpenAI Pando open source?
No. OpenAI says it is exploring ways to share the technology, but there is no public release yet.
Does Pando replace Git?
No. It replaces the server side only. Clients and the Git protocol stay the same, which is why it’s called Git-compatible.
Who wrote Pando?
According to The New Stack, one engineer designed the core and Codex agents implemented it in about 10 weeks, using Rust.
Pando is a good example of AI tools forcing the rest of the stack to catch up. If you only use Git casually, you can skip the details. If your team is adding agents fast, treat your Git server and CI as part of the rollout, and measure them now. OpenAI had to build its own backend, but most teams can get far by fixing queues and branch habits first. For another side of this shift, see how GitHub Copilot Fusion picks the model for you.
Leave a Reply