On November 2, 2026, OpenAI engineers Aaron Friel, Taylor Blau and Ted Nyman published "Scaling Git at OpenAI", describing Pando, "a new backend for Git repositories" that "now powers all development in OpenAI's monorepo". Two days later OpenAI announced it would acquire Pierre Computer Company, and that announcement links the two: "Earlier this week, we shared how we're scaling Git at OpenAI with Pando, our in-house source control backend for engineers and agents." The Pierre post also says Pierre's open-source projects "will remain open-source", that Code Storage customers can keep using the service after closing, and that the deal is subject to regulatory approvals.
Why the old setup broke
The post blames parallel work by coding agents. "One engineer now has several changes ready for CI at once, and each push can set several jobs fetching the same commit." From around the end of 2025 the monorepo suffered slow fetches and pushes and failed merges, "with reliability at times falling below 90%". The team modelled two standard designs. Three-phase commit across full replicas puts quorum disk writes and round trips on every push. A shared write-ahead log in blob storage lets packfiles upload in parallel, but every push, even to an unrelated branch, still competes to update one manifest.
Two assumptions Pando drops
Pando keeps the Git protocol and client and changes two things most Git servers treat as requirements. The first is a consistent snapshot of all refs. Pando serialises updates to each ref but offers only causal consistency across the repository, so "Clients only conflict when pushing to the same ref simultaneously", and a client reading several branches "cannot assume those reads describe a single point in time across the repository." The authors argue that "users and agents are agnostic to the global ordering of ref updates".
The second is a custom packfile for every fetch. Each push is stored as an immutable packfile in blob storage, with a ref version recorded in a metadata store. When pushing a new branch, Pando advertises only the trunk, so the pack depends on nothing else and can be handed unchanged to later readers. A full ref advertisement with millions of branches "can exceed 100 MiB". The cost is that pushes resend everything outside the trunk; in the monorepo the median pack was under 20 KiB and the 99th percentile under 3 MiB. Fetches return packfile URIs, a less-used Git protocol feature, and a cache service called Express serves them with hit rates above 99 percent. Background compaction groups small packs into larger covers, up to level L4 today, much like a log-structured merge tree, so fetch planning takes logarithmic rather than linear work.
Reported results and how it was built
OpenAI reports that switching "reduced the latency of the median push by nearly 80%", that Pando "can deliver a petabyte of Git data per day with just a handful of CPU cores", and that it once served more than 57 terabytes of packs in an hour, over 99 percent from local disk cache. Clones on standard cloud VMs saturated 10 Gbps links. Pando is written in Rust and went from kick-off to production in 10 weeks, designed by one engineer and "implemented by teams of agents", with a Git conformance suite, deterministic simulation testing, and Lean and TLA+ models written with agent help. It began as a read- and write-through proxy in front of the old system, so traffic moved without downtime. Several client changes, including parallel pack downloads and "explicit haves", are in OpenAI's public Git fork, and some are already upstream.
What remains uncertain. All production figures are OpenAI's own, from one very large repository. The interactive comparison in the post is labelled "Synthetic teaching models, not production benchmarks". Pando is not available outside OpenAI; the post says only that OpenAI is "exploring ways to share the technology". How it behaves for tools that do need a cross-branch snapshot is not discussed in detail.
Analysis: The transferable idea is to ask which consistency a workload really needs before scaling the store. At Belgavi's Book Store, Jenya's help-desk team lets agents push draft reply templates to one branch per CASE. Per-branch ordering is enough for that. The step to check is the weekly release job, which reads several branches and merges the approved ones: anything that assumes all branches were read at the same moment should instead record the exact commit it read from each branch. Writing those commit IDs into the release notes costs nothing on any Git host and keeps the job correct if the backend ever relaxes its guarantees.
Sources and evidence
OpenAI's Pando (described November 2) is a Rust, Git-compatible backend now serving its whole monorepo. It orders updates per branch rather than per repository, stores each push as a reusable packfile, serves packs by URI from regional caches and compacts them in levels. OpenAI reports median push latency cut by nearly 80 percent and about a petabyte served per day; the figures are its own, and Pando is not publicly available.