Daily Issue #11 · 7 min read

Coding agents broke Git at OpenAI. Now it is buying the tools to rebuild it.

OpenAI's deal for Pierre Computer, two days after it revealed a new Git backend called Pando, shows where the coding agent race is heading: code review and source control.

The short answer

On November 4, OpenAI said it will acquire Pierre Computer Company, a small developer tools startup that makes Code Storage, a Git based service for storing code and its history, plus open source libraries called Trees, Diffs and Highlights. Trees and Diffs already power file browsing and code review inside the Codex app. Financial terms were not disclosed and the deal still needs regulatory approval. Two days earlier, OpenAI described Pando, a new backend it built because its own Git setup could not keep up with the volume of work Codex agents were creating. Put together, I read this as OpenAI moving past writing code and into the place where code is stored, reviewed and shipped. That is the territory GitHub has owned for well over a decade.

What OpenAI is actually buying

Pierre is not a model company, and that is the point. According to OpenAI's acquisition announcement, the deal brings four things into Codex:

• Trees, a library for quickly browsing a project's files.
• Diffs, a library for seeing what changed, with an Edit mode that lets a reviewer fix a line right where they spot the problem.
• Highlights, a syntax highlighter fast enough to keep up while a developer types or an agent streams code.
• Code Storage, a service that lets apps store code and track its history using Git.

OpenAI says Pierre's open source projects will stay open source and that existing Code Storage customers can keep using it while its features move into Codex. The stated ambition is bigger than a product feature. OpenAI wants to "advance the future of source control across the industry" and to make Codex the place where work goes "from the first branch an agent creates to that magic moment when code is deployed to the world."

Why Git became the bottleneck

The more interesting document is the one OpenAI published on November 2, titled "Scaling Git at OpenAI." It explains the problem in plain terms. Codex lets each engineer run several changes at once, so there are far more branches, pushes, merges and fetches than a human team would create. Each push can also kick off several CI jobs that all fetch the same commit.

OpenAI says that starting around the end of 2025 its monorepo systems were struggling, with slow fetches and pushes, failed merges, and reliability that at times fell below 90 percent. Its answer was Pando, a new backend that still speaks normal Git, so engineers and agents did not have to change how they work. The main design change is that updates to different branches no longer wait on each other. The tradeoff, which OpenAI states openly, is that someone reading several branches at once may see them at slightly different points in time.

The results OpenAI reports are large:

• Median push latency fell by nearly 80 percent, and merges became more reliable.
• Pando delivers nearly a petabyte of Git data per day and once served more than 57 terabytes in a single hour.
• It went from kickoff to production in 10 weeks. OpenAI says one engineer designed the core, teams of agents wrote it, and a handful of engineers deployed it.

OpenAI also says it is exploring ways to share the technology behind Pando with its users. Read next to the Pierre deal, that sentence matters more than the benchmarks.

The real prize is code review

Storage is the plumbing. Review is where the pain is. Once agents can write code faster than people can read it, the slow step in shipping software is a human looking at a diff and deciding whether to trust it. That is exactly the part of the workflow Pierre's libraries were built for.

I think this is the right bet. A coding agent that produces ten pull requests an hour is only useful if someone can review ten pull requests an hour. Better models help a little, because cleaner code is easier to check. Better review tools help a lot more: fast file browsing, diffs that stay quick on huge changes, and the ability to fix a small mistake without bouncing the whole change back to the agent.

OpenAI's own wording points the same way. It wants Codex to "connect the code a company ships with the decisions its people make." I read that as saying the record of why a change happened, who approved it and what the agent was told becomes part of the product. Whoever holds that record holds a lot of power over how a team builds software.

Where this leaves GitHub

Axios, which had the deal as an exclusive, framed it as OpenAI taking aim at Microsoft's GitHub. Axios also noted that GitHub has had reliability problems as AI coding has grown, and that critics say it was built for humans at a time when more and more code is written by machines. OpenAI is not alone here. Axios points to Origin, a GitHub rival that Cursor launched in August, and to former GitHub CEO Thomas Dohmke, whose new coding startup announced a competing product last month.

There is an awkward twist. Axios notes that Microsoft owns nearly a quarter of OpenAI. So one of GitHub's biggest backers is also funding a company that now wants to rebuild the layer GitHub sits on. That is common in AI right now, but it is still worth noticing.

I don't think GitHub is about to lose its place overnight. Its strength is the network: open source projects, issues, permissions, integrations and years of habit. Pando itself shows how much OpenAI values that habit, since the whole design keeps the standard Git client working unchanged. The likelier near term change is quieter. More code will be written, reviewed and merged inside the agent tool, and the hosting platform may become one stop on the way rather than the center.

What to watch before you get excited

• The deal has not closed. OpenAI says it depends on customary conditions including regulatory approval, and until then the two companies stay independent.
• No price was disclosed, so it is hard to judge how big a bet this is in money terms.
• Pando's numbers are OpenAI's own, measured on OpenAI's own monorepo. A team with a few hundred branches will not hit these limits any time soon.
• The bigger question is lock in. Open source libraries staying open is good news, but where your code history, review comments and approvals end up living is a much bigger decision than which model you use, and it is harder to undo.

What this means if you ship software with agents

If your team has started running several coding agents in parallel, I would look hard at the review step before buying more agent capacity. Count how many agent pull requests sit waiting each day and how long a human takes to clear one. If that queue keeps growing, a stronger model will not fix it. Smaller, well scoped changes, clear rules for what an agent may merge on its own, and review tools that stay fast on large diffs will do more.

I would also keep source control choices boring for now. Standard Git gives you the freedom to move between hosts and agent tools as this market settles. OpenAI seems to agree, since it rebuilt its whole backend without asking anyone to stop using normal Git commands.

My takeaway

For two years the coding agent race was about which model writes the best code. This week OpenAI showed where it thinks the next fight is: the systems that store code, show what changed and let a person say yes. Agents made Git slow enough at OpenAI that it rebuilt the backend in 10 weeks, and now it is buying the team behind its review tools. When a company running coding agents at that scale decides the bottleneck is review and source control, I take that as a strong signal. The next big gains in AI coding may come less from smarter models and more from helping people trust what those models produce.

Quick notes

• Apple's annual report warns AI agents may resist instructions: In its annual report filed with the SEC on November 3, Apple added a risk factor saying AI agents, its own or third parties', may take actions that were not authorized, including exceeding permissions, defying directions or resisting instructions. Apple also wrote that its ability to oversee third party agents may be limited and that some consequences, such as deleted data or completed transactions, may be difficult or impossible to reverse.
• ChatGPT can now make GIFs with Images 2.6: OpenAI released Images 2.6 on November 4 across ChatGPT, Codex and the API. It adds animated GIFs and SVG export, and OpenAI says it generates images 50 percent faster. Developers get two API models, GPT Image 2.6 Flare for speed and Sunburst for detail. OpenAI says people now create 4 billion images a week across its products.

Common questions

What is OpenAI acquiring with Pierre Computer Company?

OpenAI is acquiring Pierre Computer Company, the maker of Code Storage, a Git based service for storing code and its history, and the open source Trees, Diffs and Highlights libraries. Trees and Diffs already power file browsing and code review in the Codex app. OpenAI announced the deal on November 4, 2026.

How much did OpenAI pay for Pierre?

OpenAI did not disclose the price. The deal still needs regulatory approval and other customary conditions, and the two companies remain independent until it closes.

What is Pando at OpenAI?

Pando is a backend for Git repositories that OpenAI built to handle the volume of branches and pushes created by engineers working with Codex agents. OpenAI says it cut median push latency by nearly 80 percent and delivers nearly a petabyte of Git data per day while keeping the standard Git client unchanged.

Will Pierre's open source tools stay open source?

Yes, according to OpenAI. It says Pierre's open source projects will remain open source, that it will keep contributing to them, and that Code Storage customers can keep using the service after the deal closes.

Is OpenAI competing with GitHub?

Axios reported the deal as OpenAI targeting GitHub, as AI agents take on more coding work. OpenAI says it wants Codex to cover work from the first branch an agent creates to deployment, which overlaps with what GitHub does today.