OpenAI has built a new backend for Git repositories called Pando, after its coding agent Codex pushed the company’s shared codebase past what its old source-control setup could handle. In a post published November 2, the company says Pando now powers all development in its monorepo, cutting median push latency by nearly five times and serving almost a petabyte of Git data every day. The system went into production in April 2026.
Why the old setup started to fail
Codex changed the shape of work inside OpenAI. One engineer can run agent sessions in parallel, each producing branches, pushes, merges and fetches, while development environments, research jobs and integration systems clone and fetch commits. By the end of last year, that load began to show: pushes and fetches slowed, merges failed, and reliability at times fell below 90 percent.
OpenAI says it asked a question before rebuilding anything: had it hit a limit of Git itself, or only of the way it served its monorepo? Its answer was the second. Git has carried software development for more than two decades, and the company wanted to keep the workflow engineers and agents knew, rebuilding only the server side underneath it.
Two ideas that make Pando different
Pando gives up two things most Git systems treat as requirements. First, it drops repository-wide consistency: updates to unrelated branches proceed independently, under what OpenAI calls causal consistency, so a reader looking at several branches can see them at different points in time. Second, it reuses packfiles across fetches instead of building a custom pack for every client.
The second idea depends on a small trick. A normal server advertises every branch, so the advertisement alone can exceed 100 megabytes with millions of branches. Pando advertises only a select set, by default a trunk branch, so a client’s pack depends only on what the server allowed it to leave out, and that pack can later serve a fetch.
That trade costs bandwidth: a client pushing the same branch repeatedly, or stacking one feature branch on another, resends everything outside the trunk. In OpenAI’s monorepo the median pushed packfile was under 20 kilobytes and the 99th percentile under 3 megabytes, which the company judged a fair trade for reusable packs and fast reads.
Built by agents, checked like a proof
OpenAI says it built Pando with Codex, moving from kickoff to production in just 10 weeks. The core was designed by a single engineer, implemented by teams of agents and deployed by a handful of engineers, in Rust, as a new codebase without legacy dependencies. As new models arrived during the build, the teams used them to improve the system.
Validation mattered as much as speed. The teams wrote a conformance test suite so agents could check changes against standard Git behavior, built deterministic simulation harnesses to hunt race conditions in negotiation and pack planning, and used GPT-6 Astra to write Lean and TLA+ proofs formalizing a theory of Pando.
Agents also red-team the system around the clock using Daybreak-capable models. ChatGPT and Codex will share one usage pool from December, a sign of how central Codex has become to OpenAI’s own work. That testing culture is part of a wider pattern at OpenAI, where models are asked to produce work a checker can verify rather than simply asserting it.
Deployment avoided a big-bang switch. Pando first ran as a read- and write-through proxy in front of the existing system, keeping both in sync while traffic moved over system by system. OpenAI shipped its own Git distribution to developer machines and inside the ChatGPT desktop app, and migrated hundreds of integrations that depended on legacy interfaces, webhooks and repository downloads.
Why It Matters
Pando is a measure of how far coding agents have already changed software work. The bottleneck was not model quality or chip supply but the plumbing underneath development: version control built for humans pushing a few times a day, now asked to absorb agents pushing constantly.
That scale is not unusual anymore. Agents already draft 40 percent of GitHub’s code changes, and every one of those changes flows through the same kind of infrastructure, which is why a backend rebuild at one company reads as a warning for the rest.
For other engineering teams, the lesson is practical rather than theoretical. If agents multiply branches and fetches, push latency, merge failures and fetch times deserve attention before uptime dashboards do. OpenAI’s answer will not transfer directly to smaller teams, but its diagnosis, that the server side, not Git itself, was the limit, gives vendors a clear direction for agent-era hosting.
Our read
The most telling choice is what OpenAI refused to change. Many companies at this scale replace Git, forcing every tool and agent to relearn version control. OpenAI kept the protocol, so models that know Git commands keep working, and its client improvements can flow upstream to Git. That restraint is also competitive: it keeps OpenAI’s agent workflow portable.
Still, every headline number here is OpenAI describing its own system. There is no independent measurement, no baseline for the old backend and no hardware specification behind the handful of CPU cores said to serve a petabyte a day today.
OpenAI says only that it is exploring ways to share the technology, with no product name, date or price, so for now Pando is evidence of a problem and one company’s answer, not a tool anyone else can buy yet.
Background
Git grew out of a community exchanging kernel patches by email, and developers have adapted their tools around it for more than two decades as the way software is built changed. Pando is the latest adaptation, aimed at a shift its builders did not plan for: a single engineer behaving, through agents, like a small team.
0 Comments