Written by Henk Boelman on November 8, 2026

Why everyone is suddenly rebuilding Git

Anthropic replaced Git, OpenAI rebuilt its engine and is buying a code hosting startup, and Entire, Cursor and GitHub are all racing to host agent-written code. We look at why AI agents broke the scale Git was built for, and what it means for your team.

6 min read

In the space of a few weeks, Anthropic revealed it replaced Git for its main codebase, OpenAI unveiled a new Git backend and announced it is buying a code hosting startup, Entire launched a forge built for agents, and GitHub said it is rebuilding its Git architecture from the ground up. Add Cursor's Origin from August, and the pattern is hard to miss.

Git has been the backbone of software development for two decades. So why is everyone rebuilding the layer underneath it now? The short answer: AI agents broke the scale it was designed for.

The problem: agents write more code than Git was built for

Git was designed for humans. A developer makes a handful of commits a day, works on a local copy and pushes changes now and then. Agents work differently. They run in parallel, open branch after branch and push around the clock.

The numbers show how fast this changed. At Anthropic, merged changes in its monorepo, a single repository holding all of a company's code, grew from about 2,500 a day in May to more than 30,000 by September. GitHub recorded 7.38 billion commits in September alone, and its Octoverse report says more than 40% of pull requests that month started with a draft from an AI agent.

When writing code stops being the bottleneck, the bottleneck moves downstream: to source control, builds and CI (continuous integration, the automated building and testing of every change). Those systems were sized for human commit rates, not swarms of agents.

Two AI labs, two opposite answers

Anthropic and OpenAI hit the same wall, but solved it in opposite ways.

Anthropic: replace Git

A team of seven at Anthropic built Scion, an in-house version control system that replaced Git for its monorepo, and completed the migration in five months. Scion is centralized and does most of its work on servers instead of on local copies. The reasoning: almost all its users are agents and automated jobs that are always online, so Git's biggest strength, working offline with a full local copy, adds little and costs a lot. Alongside Scion, Anthropic built its own code review tool, code search and access control for source code.

Anthropic is clear that Scion is built for its own situation, not as a Git replacement for everyone, and it has no plans to open source it.

OpenAI: keep Git, rebuild the engine

OpenAI took the opposite route. In Scaling Git at OpenAI, published on 2 November, it describes Pando: a new backend that serves all development in its monorepo while engineers and Codex agents keep using the normal git push and git fetch commands.

The trouble started around the end of 2025. As Codex let engineers run many agents at once, branches, pushes and merges multiplied, and the reliability of the old setup at times fell below 90%. OpenAI concluded the problem wasn't Git itself, but how Git servers are traditionally built. Pando changes two things:

Branches move independently. Instead of coordinating every update across the whole repository, Pando handles each branch on its own, so pushes to different branches no longer wait in line.
Downloads are reused, not rebuilt. Normally a Git server builds a fresh packfile, Git's compressed bundle of changes, for every client that asks. When dozens of agents and CI jobs request the same code, Pando serves the same stored packs from a cache instead.

The results, according to OpenAI: median push latency down by nearly 80%, nearly a petabyte of Git data served per day, and more than 99% of data coming from cache. Pando has been in production since April, and OpenAI says it built it in Rust with Codex, going from kickoff to production in ten weeks. It is internal only, but OpenAI says it contributes improvements back to Git itself.

From plumbing to products

Two days after the Pando post, OpenAI announced it will acquire Pierre Computer Company, the startup behind Code Storage, a Git hosting service built for machine traffic. Pierre's open source Trees and Diffs libraries already power file browsing and code review in the Codex app. OpenAI plans to bring Code Storage into Codex, with the stated goal of making Codex the place where developers and agents do their most ambitious work. The deal still needs regulatory approval.

That puts OpenAI next to the companies building code hosting for everyone else:

Entire launched a forge, meaning hosting, review and builds in one service, with AI-scored reviews instead of pull requests and regional hosting built for agent traffic.
Cursor launched its own hosting service, Origin, in August.
GitHub, the incumbent, announced at Universe that it added 4 million CPU cores this year, is moving to a multi-region setup on Azure and is rebuilding its Git architecture to separate storage from compute.

Why now?

Three forces come together.

Scale. This isn't entirely new. Google and Meta outgrew standard Git years ago and built their own systems, Piper and Sapling, and Anthropic built Scion on that prior art. What's new is that agents give a single AI lab, or soon any software company, the commit volume of a tech giant. Many more organizations are hitting the same wall at once.

Reliability. When agents ship most of your code, an outage at your code host stops the whole company. OpenAI's interest in its own platform reportedly grew after GitHub outages left its engineers unable to work earlier this year. Owning the stack becomes a way to own your uptime.

Control of the workflow. Code hosting is no longer just storage. It's where agent work gets reviewed, tested, remembered and approved. Whoever owns that layer sits at the center of how software gets built, which makes it strategic territory for AI labs, tool makers and GitHub alike.

What this means for you

Should your team drop Git? Almost certainly not. Even Anthropic says Scion is built for its specific situation, and OpenAI went out of its way to stay fully compatible with Git. For most teams, Git and their current host will keep working fine.

What does change is everything around it:

Review is the new bottleneck. If agents produce more pull requests than your team can read, the answer is better review: automated checks, clear rules for when a human must look, and smaller changes.
Your CI was sized for humans. If builds and tests already queue up, agents will make that worse. Measure it now, before you scale up agents.
Watch the platforms. Entire, Origin, GitHub and now Codex are all competing to become the place where agent work lives. The ideas they test, from scored reviews to agent memory, will likely become standard.

The deeper lesson is the one we keep coming back to in our community: when code gets cheap to write, the value moves to judging it. The companies rebuilding Git are simply the first to feel that at full scale.

How is your team handling the extra volume from agents? Share your experience in the community channels or bring it to your next chapter meetup.