Anthropic Scion: version control built for AI agents, and why Git couldn't keep up

Anthropic Scion is the version control system Anthropic built in-house after Git stopped keeping up with its AI agents. In October 2026 the company described how it moved its monorepo to Scion: from about 2,500 merges a day in May to more than 30,000 in September. It is not a general replacement for Git, it is internal, and every figure comes from Anthropic itself.

What Anthropic Scion actually is

Almost every headline will say "Anthropic replaced Git." That is not what the post says. Anthropic wrote that Scion is not a general replacement for Git. The real news is that, in its words, writing code is no longer the bottleneck. Version control, builds, and continuous integration are, because all of them were designed for human speed.

The migration of its monorepo, the giant repository where the code lives, took five months and a team of seven people.

Why Git breaks when the clients are agents

Git is decentralized: every clone is a full repository that works offline. Anthropic, however, was running it as a centralized system, with almost every client being an agent and jobs that are always online. The properties built for working without internet became dead weight.

The team tried shallow clones, partial checkouts, and internal mirrors. At the peak it operated several hundred Git mirrors. The mirrors drifted behind the main branch between updates, and diff and blame on partial clones got harder and harder to sustain.

The six ideas behind Scion

1. Clients are always online, so they request what they need instead of copying everything.
2. The server does the heavy work: history, blame, commits, merges, and rebases.
3. Version control is a service with an API, so many jobs have no local repository.
4. Everything scales horizontally, and not even the server holds a full copy of the repo.
5. Commits happen on the server, so there is nothing to push.
6. Checkout is optional, partial, and virtual: files appear when you touch them.

The life of a change, step by step

An agent asks for a checkout of only the directories it needs. The request reaches the scion-daemon on the same node, which stores objects for every job on that machine. If the object is missing, it asks the regional scion-cache, which serves it from memory, and after that the Scion servers, which read from Spanner or cloud storage.

When the agent submits its change, the server creates a changeset: a list of files modified relative to the first parent, without building full trees. Nothing is pushed. In parallel, Scion computes derived data such as blame, per-path history, and shortcuts for walking the commit graph, so a new type can be added without touching the clients.

Moving the main branch is a single compare-and-swap operation on one Spanner row. Two concurrent changes can collide, and Anthropic says Claude usually fixes the conflict faster than a human can spot it. Instead of a global lock, it relies on post-merge tests.

The numbers Anthropic reports

• Twenty-five million CI checkouts a day at 25 GB of compressed working copy each: 625 petabytes a day if pulled in full, which Anthropic rounds to over 600.
• From 2,500 merges a day in May to more than 30,000 in September: on average one merge every 35 seconds against one every three seconds.
• Infrastructure 200 times smaller in CPU and memory, checkouts that go from minutes to seconds, and 25 percent more CPU capacity in CI alone.
• Every cache tier hits more than 99 percent of the time.

All of these are self-reported by the team that built the system, on its own blog, with no outside audit. The team also worked with an early internal snapshot of Claude Mythos Preview and admits its token spend far exceeds what another company would spend. The 200x is measured against its own mirror setup, not against a well-configured Git. Treat the numbers as a direction, not a promise.

Who Scion is not for

If your repository fits comfortably on a laptop, Scion is a rocket for going to the corner store. Anthropic says its ideas only make sense for huge monorepos with tens of thousands of engineers or swarms of agents. And you cannot download it: Scion is internal, with no license and no public version.

What actually breaks

According to the post, agents take shortcuts: several systems scanned the entire history to find when a change was merged instead of asking the review tool. Adversarial review was sometimes too defensive and added extra validation and complexity. A global lock on the merge queue does not hold up when hundreds of changes merge while tests run. And replacing Git means replacing code review, code search, and continuous integration too. To catch subtle failures, Anthropic uses Claude Tag instances, one per component, watching the daemon, the cache, and the merge queue in Slack channels.

How to apply it in your business?

You do not need Scion to use the lessons. These are the ideas to copy:

1. Measure. Count how many checkouts your CI does a day and how many gigabytes each pulls; that multiplication is your petabyte bill.
2. Clone less. In CI, use shallow clones and partial checkouts of only what each job needs.
3. Put a cache close to your runners so not everyone hits the central server.
4. Ask, don't clone. Use your platform's API for history questions.
5. Skip the global lock. Bet on post-merge tests and automatic rollback.

The warning that prevents the costliest mistake: don't migrate off Git just because Anthropic did. It has a dedicated team of seven engineers; you probably only need two of these five. If agents are already multiplying your pull requests, read the full breakdown of how Claude managed agents handle dynamic workflows and their limits.

FAQ

Did Anthropic replace Git with Scion?

No. Anthropic migrated its own monorepo to Scion, but it states that Scion is not a general replacement for Git.

Can I download or use Anthropic Scion?

No. Scion is internal: there is no public version and no license.

How many merges a day does Anthropic handle with Scion?

It reports going from about 2,500 merges a day in May to more than 30,000 in September, roughly one every three seconds.

Are the performance numbers independently verified?

No. They are self-reported by Anthropic on its own blog, with no outside audit, and the 200x comparison is against its previous mirror setup.

Conclusion

What matters about Anthropic Scion is not that the company left Git. It is that agents move the bottleneck from code to version control. The open question is how long it takes to reach your team. How many minutes of continuous integration does a single change use on yours today?