OpenAI's own engineers couldn't push code fast enough on plain Git once Codex agents joined the commit queue, so the company spent ten weeks building a replacement.
Quarterly commits inside OpenAI's internal monorepo jumped 2,442% between the first quarter of 2024 and the fourth quarter of 2026, according to OpenAI's own engineering blog post published November 2, 2026. That is not a typo-sized number. It's the kind of growth that happens when a company's software engineers stop being the only ones writing code, and autonomous Codex agents start pushing commits alongside them around the clock.
The company's answer is called Pando, a Git backend built from scratch in Rust and now running underneath OpenAI's entire monorepo. It replaced the server-side plumbing while keeping the Git commands engineers and agents already type exactly the same. Nothing about the workflow changed. Everything underneath it did.
Git was built for a world where a finite number of humans push code during business hours and merge conflicts are rare enough to resolve by hand. That assumption breaks down when dozens of Codex agents are running in parallel, each opening branches, each pushing updates, each expecting the repository to stay fast and consistent no matter how many other pushes are landing in the same second. By late 2025, OpenAI's existing monorepo systems were straining under exactly that load, and traditional Git hosting, built around one global ordering of changes, doesn't scale horizontally the way that kind of concurrent, agent-driven traffic demands.
Pando's fix is a shift to causal consistency. Instead of imposing one global sequence on every change across the whole repository, it coordinates updates independently per branch, so unrelated branches can accept pushes at the same time without waiting on each other. That is the detail that actually matters here: it's not a faster version of the old architecture, it's a different model of what "consistent" even means for a repository this size.
OpenAI turns Codex safety into infrastructure OpenAI's Codex safety approach shows why coding agents are becoming an infrastructure problem, not just a model quality problem. Startups can copy the core ideas by using isolated environments, tighter permissions, visible approval loops and better monitoring. - how to safely deploy autonomous coding agents - building infrastructure around AI code generation models
The numbers OpenAI reported are the kind that back up the architectural bet. Median Git push latency dropped by nearly 80%. Merge reliability, which the company says had sometimes dipped below 90% under the old setup, now sits at 99.99%. Pando serves close to one petabyte of Git data a day, and in testing it pushed through more than 57 terabytes of packfiles in a single hour, enough to saturate a 10-gigabit network connection on clone and fetch operations. Cache hit rates run above 99%.
What makes the build itself notable is who did the work. A single engineer designed Pando's core architecture, and teams of Codex agents handled the implementation, with a small group of engineers deploying it, according to OpenAI. Ten weeks, start to production. The company's developer team put it plainly in the post announcing the system: they built Pando "to handle more pushes and fetches without changing how our engineers and agents use Git."
Pando entered production in April 2026 and has been carrying all monorepo traffic since, including the hundreds of internal integrations, CI pipelines, and research jobs that depend on it, without anyone needing to change a Git command. That migration detail is easy to skim past, but it's the actual engineering achievement: you can rebuild the entire backend of a system this central and have nobody downstream notice until you tell them.
This lands in the middle of a broader scramble across OpenAI's engineering org to rethink source control for agents rather than for humans alone. A public OpenAI fork of Git appeared on GitHub back on July 10, 2026, and Git contributor Taylor Blau, formerly of GitHub, is now working inside OpenAI on source code management tools, as OpenAI's own hiring announcements and multiple outlets including Interesting Engineering have noted. Separately, OpenAI shipped Symphony in April, an open-source specification for routing Codex agents against tickets in tools like Jira and Linear, and a Codex app for macOS that runs multiple agents in parallel using Git worktrees so they don't trample each other's changes.
None of those tools replace what Pando does. Symphony coordinates what agents work on. Worktrees keep agents from colliding inside a single checkout. Pando is further down the stack: it's the thing that has to stay standing no matter how many of either are running at once. OpenAI says it's now looking at where else large numbers of software agents might need to collaborate on shared repositories, which suggests Pando was never really a one-off infrastructure fix. It's a bet that the choke point in AI-assisted software development is shifting from how fast a model can write code to whether the plumbing underneath can survive that many hands, human and otherwise, touching the same repository at once.
Also read: Sarvam AI's Prism model answers in 44 milliseconds instead of writing text • Lovable AI Apps Left Passport and Payment Data Open to Anyone • Revolut Shows Cheap Unlimited Tokens, Not Smarter AI, Is What Speeds Up Coding
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.