OpenAI's Pando: a Git backend that drops repo-wide consistency so Codex agents can push in parallel, cutting median push latency about 5x

OpenAI has replaced the backend behind its monorepo with Pando, a new Git server built so that thousands of engineers and Codex agents can push to different branches at the same time. In Scaling Git at OpenAI, Aaron Friel, Taylor Blau and Ted Nyman write that Pando cut median push latency by nearly 5x (about 80%) and now serves nearly a petabyte of Git data a day. Clients still run ordinary Git.

The reason is agents. "Codex has changed the scale and pace of development at OpenAI," the post says: more branches, pushes, merges and CI fetches, with one engineer often having several changes in CI at once. From around the end of 2025 the old system struggled, with slow fetches, failed merges and reliability "at times falling below 90%." The New Stack reports OpenAI saying quarterly commits to the monorepo will be up more than 2,400% by year-end compared with Q1 2024.

The two things Pando gives up

OpenAI's argument is that Git itself wasn't the limit; the way Git servers are usually built was. Pando drops two assumptions most of them make.

Repository-wide consistency. Conventional replicated Git servers order all ref updates in a repository together, whether through three-phase commit across replicas or a shared write-ahead log, so a push to one agent's branch can wait on another's. Pando coordinates each branch independently and offers causal consistency instead. Pushes only conflict when they hit the same ref. The trade-off, which OpenAI states plainly, is that a client reading several branches may see them at different points in time. OpenAI's view is that "users and agents are agnostic to the global ordering of ref updates."

A custom packfile for every fetch. Normally the server computes which objects a client is missing and compresses a fresh pack for it, which is expensive and hard to cache. Pando stores each pushed packfile as an immutable blob, records the branch tip and its dependencies in a metadata store, and on fetch hands back links to packs it already has through Git's packfile-uri protocol feature. A caching service called Express serves them from regional caches with a hit rate above 99%, and clones and fetches can saturate a 10 Gbps link on standard cloud VMs.

To make pushed packs reusable, Pando advertises only the trunk branch to a pushing client rather than all refs. With millions of branches, a full ref list can exceed 100 MiB. The cost is that repeated or stacked branch pushes resend everything outside trunk, though OpenAI says the median pack in its monorepo is under 20 KiB and the 99th percentile under 3 MiB. Background compaction, modelled on log-structured merge trees, merges small packs into larger "covers" (up to L4 today), so a client that is far behind can be served in logarithmic time instead of downloading tens of thousands of packs.

Built with Codex, checked with Lean

Pando is written in Rust. OpenAI says it went from kick-off to production in 10 weeks, with the core designed by one engineer, implemented by "teams of agents" and deployed by a handful of people. It has served production traffic since April. Agents tested changes against a Git conformance suite and deterministic simulation harnesses. Using GPT-6 Astra, they also wrote Lean and TLA+ specifications, a "theory of Pando", that were used to find bugs and prove properties of the system. Migration ran first as a read- and write-through proxy in front of the old system, with Codex skills that moved developers' checkouts over.

What's open and what isn't

Client-side fixes go to Git. OpenAI says it fixed concurrent, interrupted and resumed packfile-uri downloads and added tracing for them, and that several of these patches are already upstream. Others, such as parallel pack download and indexing and an "explicit haves" feature for --force-with-lease pushes against trunk, so far live in OpenAI's public Git fork and patch tracker. OpenAI's Git build now ships to developer machines and inside the ChatGPT desktop app.

Pando itself is not open source. OpenAI says only that it is "exploring ways to share the technology" with users. The New Stack points out the contrast with Anthropic, whose own agent-scale source-control system, Scion, the outlet reported on in October: Anthropic built a new client and architecture rather than staying on Git, and has no plans to release it either.

The main figures (5x, a petabyte a day, 57 TB served in a single hour) are OpenAI's own production numbers. The interactive comparison with three-phase commit and WAL designs in the post is labelled "synthetic teaching models, not production benchmarks."