OpenAI Builds Its Own Git Backend in 10 Weeks as Codex Drives a Surge in Code Commits

The system, named Pando, went from kickoff to production in 10 weeks, according to the post and coverage by Interesting Engineering and AI News Weekly. One engineer designed the core architecture, teams of Codex agents wrote the implementation, and a small group of engineers deployed it. Pando is written in Rust and has been running in production since April 2026.

The Scale Problem

OpenAI said commits to the master branch of its monorepo rose 1,378 percent between the first quarter of 2024 and the third quarter of 2026, and it estimates the increase will reach 2,442 percent by the fourth quarter of 2026, according to AI News Weekly's account of the post. Interesting Engineering reported the 2,442 percent figure as the growth in quarterly commits over the same period. The company did not publish absolute commit numbers.

The growth came largely from Codex, which allows many engineers and autonomous agents to work on separate changes at the same time. Under the old setup, pushes and downloads slowed, merges failed, and many automated jobs fetched the same commits repeatedly. OpenAI said merge reliability sometimes fell below 90 percent.

What Pando Changes

Pando keeps the standard Git protocol and command-line client, so developers do not need new tools. The changes are on the server. Rather than enforcing a single order of updates across the whole repository, Pando coordinates updates branch by branch. Pushes to different branches can be committed in parallel, while pushes to the same branch are handled one at a time. The trade-off, OpenAI acknowledged, is that a client reading several branches cannot assume it is seeing them all at the same moment.

Incoming data is stored in cloud object storage, with a metadata service tracking each branch version. A delivery layer keeps packfiles, the bundles Git uses to move data, in regional caches and reuses them across clients instead of generating a new one for every request. Background workers merge small packfiles into larger ones.

OpenAI said median push latency fell by nearly 80 percent, merge reliability now stands at 99.99 percent, and the system serves close to a petabyte of Git data a day. At one peak it served more than 57 terabytes in a single hour, with over 99 percent of the bytes coming from local disk cache.

Why It Matters

The post gives a rare look at how AI coding agents are changing the workload inside a leading AI lab. Tools built for human-paced development are coming under strain as agents produce far more code changes, and OpenAI's answer was to have those same agents build the replacement. For engineering leaders rolling out coding agents, the figures point to source control, continuous integration and code review as likely bottlenecks.

The performance numbers are self-reported and have not been independently verified. Pando is internal: OpenAI has not released source code or offered it as a product, though the company said it is contributing many of its Git fixes back to the open-source Git project and plans to explore other settings where many software agents work on shared repositories.