OpenAI buys Pierre to rebuild Git for teams of coding agents
Two acquisitions landed on the same Wednesday, and they point at the same layer of the stack: the unglamorous machinery between a model writing code and a reviewer merging it. OpenAI went first and announced it; Meta matched it within hours.
OpenAI said on November 4 that it will acquire Pierre Computer Company, Inc., the small team behind Code Storage and a family of open-source libraries that already draw the code inside its Codex app — a deal about the one part of software work that agents have made hardest, which is reading what they wrote. If you have browsed a project's files in Codex or reviewed a change there, you have used Pierre's software: Trees for file navigation, Diffs for rendering changes, and Highlights, a syntax highlighter fast enough to keep up with an agent streaming tokens into a file, according to OpenAI's announcement.
The details matter more than the press release. Diffs includes an edit mode that lets a reviewer fix a line inside the diff instead of opening the file, which sounds like a small convenience and is not: it is the difference between noticing a bad line and actually correcting it inside the review flow. Code Storage, the company's commercial product, is a service that lets an app store code and track its history through Git, with an API Pierre designed. OpenAI says Code Storage customers can keep using it after the deal closes while the features are folded into Codex. No purchase price was disclosed. The deal still needs regulatory approval, and the two companies stay separate until it clears.
Why buy this now? Because OpenAI's own engineers published the answer two days earlier, in a post titled Scaling Git at OpenAI: Pando, an in-house source-control backend built for humans and agents at once. Git was designed around people committing in sequence, on one machine, in one branch at a time. An agent fleet does none of that. It opens dozens of branches, rewrites files in parallel, produces diffs faster than any human reviewer can read them, and leaves conflicts that are structural rather than textual. OpenAI's framing in the Pierre post is that the systems around Git have to evolve as more work happens in parallel — which is an admission that the bottleneck moved from generation to review.
The strategic read is that OpenAI just bought the last human surface in the coding loop. Code generation is a model problem, and OpenAI already sells the model. What it did not own was the place where a person looks at what the agent did and decides whether it ships — and that surface is not a niche any more, since OpenAI opened its Codex agent harness to every developer through its Agents API. Buy the diff viewer, the file tree and the review layer, and that decision happens inside Codex rather than inside a competitor's editor or inside GitHub, which belongs to Microsoft. The company is explicit about the ambition: it wants to "set the standard" for source control in the agent era, in its own words, and to connect "the code a company ships with the decisions its people make."
There is a real counter-argument, and it is the reason not to overread this. Pierre is a small team with a set of libraries that are, and will remain, open source — OpenAI committed to that, which means it cannot take Trees or Diffs away from anyone. Code review is not a moat; GitHub, GitLab and every serious IDE have shipped agent-aware review this year, and Microsoft has the distribution. What OpenAI actually bought is default placement plus a design team whose work it already depended on. That is an acquihire with a strategic story attached, and the strategic story is the part that will be tested when the libraries stay downloadable for free.
Meta made the matching move the same day, buying Conductor, the app that lets a developer run a team of coding agents in parallel — and the way it wound the product down is more instructive than the deal itself. Conductor launched in July 2025 and grew to more than 100,000 builders, per the company's own site, running first-party Claude Code, Codex, Cursor and OpenCode agents inside isolated cloud sandboxes. The founder, Charlie Holtz, confirmed the team is joining Meta Superintelligence Labs, as Axios first reported. The terms were not disclosed.
What followed is a masterclass in what "acquired" means for users. New sign-ups closed the day of the announcement. Conductor Cloud goes read-only on November 13, when active subscribers get refunds. On November 20 the service shuts down and user data starts being deleted. The local Mac app keeps receiving new models and critical fixes until early 2027, and then it winds down too. Holtz's note promises "something new" for early next year and says nothing about what it is, in the announcement on Conductor's site. Sixteen days from announcement to data deletion is fast even by acquihire standards, and every team that moved its workflow onto cloud workspaces this year is now exporting them back to a laptop.
Put the two deals together and the pattern is not subtle. In one afternoon, the two best-funded labs each bought a piece of the independent tooling layer that developers had assembled around coding agents — one the review surface, one the orchestration surface. The model vendors are not content to supply the intelligence and let a market of small tools sit on top of it. They are buying the workspace, and the founders are selling, because the alternative is competing with a distributor that also owns the model.
Nothing will ship its first AI-native product in December, and its chief executive used the announcement to argue that everyone else in consumer AI hardware is chasing the wrong thing. Carl Pei published a long essay titled "Toolmakers" on November 4, alongside a teaser image, saying the device "has no new screen, and it couldn't have been an app," in the post itself. He does not name it, price it, or say what it does.
The argument is the interesting part. Pei calls the industry's hunt for "the iPhone of AI" — the pendant, the pin, the glasses, the companion — "a dead end," on the grounds that consumer demand for AI is small next to demand from knowledge work, and that coding alone accounts for more than half of what companies spend on task-specific AI. His claim is that the missing piece is context: the model knows nothing about your day, and the tools you carry are where that context lives. The Verge notes the tease comes a little over a year after Pei first announced AI-native devices and an AI operating system, and follows reports that Nothing has been preparing smart glasses and other AI accessories.
Treat the ship date as the news and the philosophy as positioning. Screen-free AI hardware has a graveyard — the Humane pin being the loudest — and a company that has sold tens of millions of phones has an advantage those startups did not: distribution and a supply chain. But a December launch with no name, no price and no form factor is a promise, not a product, and the essay's own logic cuts against the anniversary-drop pattern it is setting up.
What to watch: whether regulators clear Pierre quickly or open a review, since that is the first real test of how the labs' tooling acquisitions are treated; whether Conductor's team surfaces a Meta-branded agent workspace before the local app dies in 2027; and what Nothing actually shows in December, because "a tool, not an app" is a claim a spec sheet can settle.
If the labs end up owning the editor, the diff and the sandbox, what is left for independent developer tools? Write to us via our contact page.
