OpenAI Pando: The Git Backend Built for Codex Agents

OpenAI Pando is the Git backend OpenAI built because its AI agents were pushing more code than Git servers could handle. On Nov 2, 2026 OpenAI's engineering team published Scaling Git at OpenAI, written by Aaron Friel, Taylor Blau and Ted Nyman. It explains how OpenAI Pando works and why it was needed.

Note the name: this is OpenAI Pando, the source-control system, not the Pando tree, the Pando app or any other Pando. Here is what it does, the numbers, and whether you can get it.

OpenAI Pando at a glance

Why OpenAI needed Pando

Codex changed how OpenAI writes code. One engineer now has several changes ready for CI at once, and each push can start several jobs that fetch the same commit. OpenAI's chart shows commits on master growing far faster than its own base-case plan, with reports putting the jump at about 2,442% from Q1 2024.

From late 2025 the old system struggled: slow fetches, slow pushes, failed merges, and reliability that sometimes fell below 90%. OpenAI had two choices. Drop Git, or keep Git and rebuild the server side. OpenAI Pando is the second choice, so engineers and agents keep using plain git push and git fetch.

How OpenAI Pando works

Big Git services usually keep every branch in one strict order. That means a push to one branch can wait on a push to a totally different branch. OpenAI Pando drops that rule.

• Each branch moves on its own. Pando uses causal consistency: updates to one ref are strictly ordered, but unrelated branches do not wait for each other. Clients only clash when they push to the same branch at the same time.
• Packfiles live in blob storage. Every push is saved as an unchangeable pack, and a metadata store records each ref version. Both scale out across machines.
• The same bytes go to many readers. Instead of building a custom pack for every client, OpenAI Pando reuses packs and serves them from regional caches, with over 99% of bytes coming from cache.
• Background compaction. Small packs (L0) are merged into bigger covers over time, up to L4 today, so a client that is far behind still gets a small download.

The trade-off is small in practice: a reader looking at many branches at once may see them at slightly different moments. OpenAI says people and agents care about the order of one branch, not the whole repo.

OpenAI Pando results

How OpenAI built Pando with Codex

OpenAI Pando went from kick-off to production in 10 weeks. One engineer designed the core, teams of Codex agents wrote much of the code, and a handful of engineers deployed it. Agents running GPT-6 Astra also wrote Lean and TLA+ specs to hunt for bugs, next to a Git conformance suite and simulation tests.

OpenAI also sends fixes back to upstream Git, for example cleaning up after an interrupted packfile download, and ships its own Git build to developer machines and the ChatGPT desktop app.

Is OpenAI Pando available, and what does it cost?

No public release, no open-source code and no price yet. OpenAI President Greg Brockman wrote on X that Pando can scale to very large monorepos, and that OpenAI is considering offering it to others. OpenAI's post adds that it is exploring other uses, including workloads where many agents work together on a shared Git repo. Treat any talk of a Pando product or pricing as a rumor until OpenAI announces it.

What OpenAI Pando means for teams using agents

The big lesson from OpenAI Pando: once agents do real work in parallel, the hard part is no longer writing code. It is keeping many workers in sync and keeping people in the loop. Most teams will never need a petabyte a day, but they do need one place where people and agents share the same context.

Yes, Sumus gives you that today. Sumus is a group chat where your team and AI teammates work in one thread. Give a brief to one AI teammate, it brings in others to work in parallel, and the results come back in the same chat for your team to review.

Sources

• OpenAI: Scaling Git at OpenAI
• Greg Brockman on X; @OpenAIDevs on X, Nov 2, 2026