Pando: OpenAI Rebuilt Its Monorepo Git Server to Keep Up With Its Agents
On November 2, OpenAI detailed Pando, a new Git-compatible backend that has served its entire monorepo since April. Designed to absorb parallel pushes from Codex, it reportedly cut median push latency by nearly 5x. It was built in ten weeks, largely by agents.
OpenAI detailed Pando on November 2, 2026, a new Git repository backend that now serves its entire monorepo. The initial driver stems directly from coding agents: with Codex, each engineer pushes multiple branches in parallel, each push triggers several continuous integration jobs fetching the same commit, and the legacy system could no longer keep pace. According to the post, reliability dropped below 90% at times starting in late 2025, plagued by sluggish fetches and failed merges.
Moving Past Two Long-Standing Git Server Assumptions
Rather than abandoning Git, the team retained the protocol and client while revisiting two assumptions most Git servers take for granted.
- No repository-wide consistent snapshot. Pando coordinates each branch independently: two pushes on different branches no longer wait on each other. Only simultaneous writes to the exact same branch conflict. The accepted trade-off: a client reading multiple branches may observe them at slightly different points in time. OpenAI argues that both humans and agents practically only depend on the update order within a single branch.
- No custom-tailored packfile for each client. Instead of recalculating an object pack for every fetch, Pando stores the packfile received during a push as-is and serves it to other clients using
packfile-uri, a rarely used Git protocol feature. To keep these packs reusable, it advertises only a single reference branch (trunk) by default to pushing clients.
Packs live in object storage, metadata resides in a distributed database, and a regional cache service dubbed Express distributes them. A background compaction process inspired by LSM-trees merges smaller packs across successive tiers (up to L4 today).
Reported Performance Figures
| Metric (according to OpenAI) | Value |
|---|---|
| Median push latency | Cut by nearly 5x |
| Git data served | Nearly a petabyte per day |
| Peak throughput | Over 57 TB served in a single hour |
| Express cache hit rate | Over 99% |
| Clone throughput on standard cloud VMs | Saturating 10 Gbps links |
Figures published by OpenAI, not independently verified.
Built in Ten Weeks, by Agents
The most revealing takeaway lies in how it was built. Pando was architected by a single engineer, implemented by teams of agents, and deployed by a handful of people in ten weeks from kickoff to production; it has been serving traffic since April 2026. Written in Rust, it relies on a Git compliance test suite to let agents validate their changes, deterministic simulation tests, and Lean and TLA+ specifications written by agents powered by GPT-6 Astra to root out bugs. The migration itself ran seamlessly through a proxy with zero downtime, while Codex assisted teams in transitioning their environments.
OpenAI states it has contributed several patches upstream to the Git project (parallel pack downloads, tracing for packfile-uri, and the concept of “explicit haves”) and is exploring ways to share Pando with its users, though no concrete announcements have been made.
Our take
Pando highlights an underappreciated byproduct of coding agents: the bottleneck is no longer just model compute, but the surrounding tooling—starting with the Git server. Opting to preserve Git compatibility rather than inventing a new tool is pragmatic, and upstream contributions stand to benefit the broader ecosystem. While the reported gains are impressive, they cannot be independently verified, and the post’s simulations are explicitly educational. For teams ramping up agent adoption, the key lesson is architectural: relax repository-wide consistency and reuse packs. Should OpenAI ultimately offer Pando to customers, it will step into direct competition with GitHub and other code-hosting platforms.
Sources
- OpenAI — Scaling Git at OpenAI (November 2, 2026) Primary source
- OpenAI Developers on X — Pando Announcement (November 3, 2026) Primary source
- The New Stack — OpenAI built itself a new Git backend, too Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
