OpenAI rebuilt its Git backend to keep up with Codex agents

The news: OpenAI on Nov. 2 described Pando, “a new backend for Git repositories” that the company says “now powers all development in OpenAI’s monorepo.” Git is the widely used system programmers use to track changes to code; a monorepo is a single shared repository that holds much of a company’s code. The post, by Aaron Friel, Taylor Blau and Ted Nyman, “Members of Technical Staff,” says OpenAI needed “to revamp our source control system to scale to the level of concurrent work that Codex makes possible.” Codex is OpenAI’s AI coding agent.

Why OpenAI says it needed it: According to the post, Codex means “more branches and pushes, more merges, and more fetches and clones from development environments, research jobs, and CI.” Starting “around the end of last year,” OpenAI says, its monorepo systems “were struggling to keep up,” with “slow fetches and pushes along with failed merges, with reliability at times falling below 90%.”

What it does: Pando keeps the standard Git protocol and client, so engineers and agents did not have to change how they work, the company says. It makes two trade-offs. First, it “coordinates updates to each branch independently,” so pushes to unrelated branches don’t wait on one another; the cost is that “a reader looking at several branches can see them at different points of time.” Second, it reuses packfiles, which the post defines as “Git’s format for compressing multiple objects together,” rather than building a custom download for each client. A client “may download some objects it already has,” but servers can send the same bytes to many clients with little computing work.

The numbers: All performance figures are OpenAI’s own production measurements. “In production, Pando cut median push latency by nearly 5x,” the post says, and elsewhere describes that as “nearly 80%.” OpenAI says Pando “can deliver a petabyte of Git data per day with just a handful of CPU cores” and “has served more than 57 terabytes of Git packs in a single hour, with more than 99% of those bytes from local disk cache.” It says clones and fetches on standard cloud virtual machines “saturate 10 Gbps network links.” An interactive comparison in the post is labeled “Synthetic teaching models, not production benchmarks.”

How it was built: OpenAI says it built Pando “with Codex, moving from kick-off to production, in just 10 weeks,” that its core “was designed by a single engineer, implemented by teams of agents,” and that it “began serving production traffic in April of this year.” It is written in Rust. The post says agents using GPT-6 Astra, an OpenAI model, wrote formal specifications in the Lean and TLA+ languages to “find and fix bugs.”

What it can’t do yet: Pando is internal. The post offers no Pando code; OpenAI says it is “currently exploring ways to share the technology that makes Pando possible with our users.” The design also has costs OpenAI names: pushing to one branch many times, or stacking branches, makes the client “resend everything outside of the trunk.”

Who it’s for: For now, OpenAI’s own engineers and agents. Git users more broadly may benefit from client fixes OpenAI says it is contributing upstream; “several of these improvements are already in upstream Git.”

Our read: The most useful part of this post isn’t the speed claims; it’s the admission that AI coding agents strained the plumbing under them. OpenAI says its source control fell to reliability “at times” below 90% as Codex multiplied branches, pushes and CI jobs. That is a cost of agents that rarely shows up in a demo: every parallel agent is another client hammering the repository, the build system and the review queue. OpenAI’s answer was to relax a guarantee most Git services treat as essential, a single consistent snapshot of every branch, betting that “users and agents are agnostic to the global ordering of ref updates.” That may hold for OpenAI’s workflow, but teams that rely on repository-wide consistency should weigh it. Readers should also keep the evidence in view: the gains are OpenAI’s own measurements, the post gives no repository size or absolute latency, and the simulation is explicitly not a benchmark. Companies piloting coding agents at scale should check whether their own Git hosting can absorb the load.

What’s next: OpenAI says it is “exploring other uses for Pando within OpenAI,” including “workloads where large numbers of agents need to collaborate on a shared Git repository,” and will keep sending changes to the Git project through a public patch tracker. It gave no date for sharing Pando outside the company.

Sources: OpenAI: Scaling Git at OpenAI (Nov. 2, 2026)