Notes from Việt Nam on law, technology, languages, history, culture and travel.

Git, Rebuilt for Agents: OpenAI Buys Pierre, Days After Showing How Codex Broke Its Own Monorepo

For most of the last two decades, “where does the code live” has been a settled question: in a Git repository, usually on GitHub. This week OpenAI made two announcements that treat that question as open again. On 4 November 2026 it said it will acquire Pierre Computer Company, Inc., a small San Francisco startup that sells Code Storage, a Git hosting service built for programs rather than people, and maintains the open-source Trees and Diffs libraries. Earlier the same week, OpenAI published a long engineering write-up on Pando, the source-control backend it built after its own monorepo started buckling under the load Codex created.

The deal itself is modest in what has been disclosed. The engineering behind it is the more interesting part.

What was announced

OpenAI’s post is short. If you have “browsed files or reviewed a change in the Codex app,” it says, you have already used Pierre’s Trees and Diffs libraries. Code Storage is described as “a service that lets apps store code and track its history using Git.” OpenAI says it wants to “make Codex the place where developers and agents take on their most ambitious projects” and, after closing, to “advance the future of source control across the industry.”

Three commitments are concrete:

  • Pierre’s open-source projects remain open source, and OpenAI will keep contributing to them.
  • Code Storage customers can keep using the service after closing “while we work to bring its features and functionality into Codex.” No end date was given, in either direction.
  • Closing is subject to customary conditions, including regulatory approvals; until then the two companies stay separate.

Financial terms were not disclosed. Axios, which had the story first, framed it as OpenAI “targeting Microsoft’s GitHub,” noting that GitHub has had a run of reliability problems as AI coding traffic grew, and that Microsoft owns roughly a quarter of OpenAI. OpenAI’s own announcement does not mention GitHub.

Pierre is not a large company. Its own changelog announces a $22 million Series A with CRV and 01A. Code Storage’s site lists Lovable, Poke, Anything, Bolt, Amp and Parahelp among its customers: app builders and agent products that create huge numbers of small repositories on the fly. Its pricing is metered like object storage rather than per seat: $0.005 per GB-hour for “hot” data touched in the last seven days, $0.0002 per GB-hour for cold data, $0.06 per GB pushed and $0.15 per GB cloned or fetched, with a 99.99% SLA, a 32 TB repository size limit and SDKs in TypeScript, Python and Go.

Pando: why OpenAI rebuilt its Git server

The acquisition makes more sense read next to the Pando post. OpenAI says that from around the end of 2025, the systems behind its monorepo “were struggling to keep up”: slow fetches and pushes, failed merges, and reliability “at times falling below 90%.” The cause was Codex. One engineer now has several changes in CI at once, each push makes several jobs fetch the same commit, and the branch count keeps growing.

OpenAI did not replace Git. It kept the Git protocol and client and rebuilt the server. The post explains the design by comparing it with two conventional ways of scaling Git:

  • Three-phase commit (3PC) across full replicas. Every replica keeps the whole repository; a ref update must be accepted by a majority. Reads are consistent everywhere, but every push pays for several coordination round trips, and replica disks sit on the write path.
  • A shared write-ahead log in blob storage. Packfiles upload in parallel, but every ref update, even to unrelated branches, contends on one manifest updated by compare-and-swap.

Pando gives up two things most Git servers treat as requirements.

Repository-wide consistency. Pando provides causal consistency: updates to a single ref are serializable, but a client reading several branches may see them at different points in time. Pushes to different branches don’t contend at all; the only contended key is the pointer for the ref being updated. The write path needs a constant number of metadata operations regardless of push size.

A freshly computed packfile per client. A normal Git server computes the difference between what a client has and wants, then compresses a custom pack. Pando stores each incoming pack as an immutable blob and, on fetch, selects previously received packs that cover what the client needs, returning them through Git’s little-used packfile-uri feature. To make packs reusable, Pando advertises only the trunk branch when a client pushes a feature branch, so the client builds its pack against trunk alone. The trade-off is redundant bytes on repeated pushes; OpenAI reports a median pack under 20 KiB and a 99th percentile under 3 MiB.

Background compaction, modelled on log-structured merge trees, merges small “L0” packs into larger covers (OpenAI’s monorepo goes up to L4), which also act as a skip list so fetch planning runs in logarithmic time. Regional caches serve packs with hit rates above 99%.

OpenAI’s numbers: median push latency cut by nearly 80% (roughly 5x), close to a petabyte of Git data served per day, and a peak of more than 57 TB in one hour. They are self-reported, from one company’s monorepo, and the post’s interactive comparison is a simulation, not a benchmark against named products.

How it was built is part of the pitch. OpenAI says Pando went from kick-off to production in ten weeks, written in Rust, “designed by a single engineer, implemented by teams of agents,” and serving production since April 2026. Agents wrote Lean and TLA+ specifications of its behaviour, which were used to find bugs. OpenAI also says it is upstreaming client changes to Git, including packfile-uri fixes, and is “exploring ways to share the technology that makes Pando possible with our users.”

Why buy Pierre, then?

Pando is a backend for one huge repository with thousands of writers. Code Storage solves the other half of agent-era source control: millions of small repositories created and thrown away by software, accessed through an API as often as through git clone. Pierre’s review tools fill the gap a backend doesn’t: Diffs renders and edits diffs (its Edit mode lets a reviewer fix a line in place), and its new Highlights syntax highlighter is built to keep up with code streamed from an agent. Put together with Codex, the plausible product is a complete loop — agent creates branch, pushes, a human reviews, code ships — that never touches a third-party host. OpenAI hasn’t said that, and no hosted Git product has been announced.

There is one technical curiosity. Pando’s write-up argues that quorum-based 3PC puts replica disks and round trips on the push path. Code Storage’s own site says it “leverages Git’s three-phase commit protocol and a ‘spoke-like’ quorum-based architecture.” The two designs are tuned for different shapes of workload, so this isn’t a contradiction, but anyone expecting Code Storage to be swapped onto Pando, or the reverse, should wait to see what OpenAI actually ships.

Open questions

  • Neutrality. Code Storage’s customers include companies that compete with OpenAI on coding agents. OpenAI promised continuity, not neutrality, and set no timeline for folding Code Storage into Codex. Customers storing generated code and “memory files” with a competitor’s subsidiary will want terms in writing.
  • Regulatory review. The deal is small, but closing explicitly depends on regulatory approval, and AI acquisitions draw more attention than they did two years ago.
  • Is Git still the right abstraction? OpenAI’s bet is yes: keep the protocol, rebuild everything behind it, and push improvements upstream. That is more conservative than tools that drop Git entirely, and arguably more useful to everyone else, but only if the upstream patches actually land.
  • The GitHub angle. Axios lists a growing field — Cursor’s Origin, a new product from former GitHub CEO Thomas Dohmke — going after GitHub’s role. OpenAI now has internal infrastructure, review tooling and a storage service. Whether it wants to host the world’s repositories, or just Codex’s, is the question this deal raises without answering.

For developers, the immediate change is nil: the deal hasn’t closed, Code Storage keeps running, and Diffs and Trees stay open source. The longer-term signal is clearer. The largest AI lab has now said, with numbers, that agent-scale coding broke a conventional Git deployment, and it is buying a company built on that assumption.

Sources