Git could no longer keep up with agents
OpenAI on Nov. 2 (Nov. 3 Japan time) unveiled on its engineering blog "Pando," a new Git backend built for the huge repository that holds its internal source code in one place. It also announced it on its developer-focused X account.
The trigger was its AI coding tool Codex. The number of tasks each engineer runs at once increased, and the number of branches, pushes, merges and fetches rose sharply. At times, the success rate of the existing system fell below 90%, the company said.
Push delays cut by about 80%
According to OpenAI, Pando cut median push latency by about 80%. It serves about 1 petabyte of Git data a day and has served more than 57 TB in a single hour. More than 99% of that comes from local storage in various locations.
Pushes to unrelated branches can now be processed concurrently. Incoming data is stored as is without being unpacked, and existing data is reused for fetches. It is written in the programming language Rust and runs on commodity hardware.
Developed with Codex in 10 weeks
Pando was built in about 10 weeks, with one engineer setting the core design and multiple Codex agents implementing it. It went into production in April 2026, placed in front of the existing system so the switch could be made without downtime.
It remains compatible with Git, so local Git software can be used as is. Some improvements have been merged into Git itself, and the rest are maintained in OpenAI's public repository. Pando itself is for internal use; the company said it is exploring ways to share the technology externally but gave no timing or form.
Changes in software development
As AI agents submit large numbers of changes at the same time, development infrastructure such as Git is also coming under strain. Uber has also disclosed "GitFarm," a Git system for huge repositories.
