Scaling Git at OpenAI

OpenAI says its source control was buckling under the weight of parallel work that Codex makes possible. Fetches and pushes slowed, merges failed, and reliability at times fell below 90%. So it built Pando, a new Git backend that now handles all development in the company's monorepo, without changing how engineers or agents use Git.

The reported results are big: median push latency down nearly 80%, close to a petabyte of Git data delivered per day, and a peak of more than 57 terabytes served in one hour, almost all of it from cache. Those are OpenAI's own figures. But the signal is clear either way. When agents multiply the number of branches and pushes, tools that worked fine for human-paced development start to fail.

We think this is one of the more honest pieces of agent-era evidence we've seen, because it isn't a demo. It's a company admitting its basic plumbing couldn't keep up with how much code its own tools now produce, and showing the engineering it took to fix it.

The central tradeoff is smart and clearly stated. Pando gives up a consistent, repository-wide snapshot of every branch at one moment. Each branch is ordered on its own, so pushes to unrelated branches stop waiting on each other. It also stops building a custom packfile for every client, and instead reuses stored packs served from regional caches. OpenAI argues people and agents care about the history of one branch, not the global order of all of them. For a monorepo with millions of branches, we find that persuasive.

The way it was built is its own story. OpenAI says one engineer designed the core, teams of agents implemented it, and it went from kickoff to production in 10 weeks, written in Rust, tested against a Git conformance suite and deterministic simulations, with agents writing formal Lean and TLA+ specifications to hunt bugs. If accurate, that's a strong case for agents doing serious systems work when wrapped in rigorous verification. The verification is the part we'd want other teams to copy.

There are caveats. Every performance number is self-reported, and the interactive comparison against other designs is labeled as a teaching model, not a benchmark. Pando itself isn't available outside OpenAI; the company says it's exploring ways to share it. Repeated or stacked branch pushes resend everything outside trunk, a cost OpenAI accepts. And the upstream Git contributions, while welcome, sit alongside OpenAI shipping its own Git build to developer machines and inside the ChatGPT desktop app.

Where this leads matters most. OpenAI expects more teams to hit these limits as engineers and agents produce more software. We agree. The bottleneck is moving from writing code to everything around it: storage, CI, review. Teams adopting coding agents should plan for that load now rather than discover it when merges start failing.

Git was designed around exchanging patches by email, and conventional scaled Git services coordinate updates through approaches like three-phase commit across replicas or a shared write-ahead log, which OpenAI modeled and found limited under heavy concurrent writes. MarkTechPost's write-up adds OpenAI's commit chart showing commits landed on the main branch up 1,378% in Q3 2026 versus Q1 2024, and notes Pando began serving production traffic in April 2026.