After merges climbed more than tenfold in four months, Anthropic stopped using Git. The company laid out the story on October 13 in an engineering post titled "Redesigning source control at Anthropic," and the numbers deserve a plain-English walk-through. Anthropic expects other companies that use AI coding tools to hit the same wall within six to 12 months. That forecast comes from a company with a stake in the story, and the mechanics are worth understanding either way.

What Git does and why it won

Git is the version control system that records every change to a codebase. Linus Torvalds created it in 2005, and it became the standard because of one design decision: every developer's computer holds a complete copy of the repository, history included. You can read the full history and commit work on an airplane. When you reconnect, you push your commits to a shared server such as GitHub or GitLab, and your teammates pull them down.

That design suits people. Developers work offline for hours and sync a few times a day.

The workload broke Git's central assumption

Anthropic's clients do not work offline. Nearly all of them are agents and automated jobs that are online by definition, and each one still needs code. The main branch history alone runs almost 100 GB, and a compressed working copy tops 25 GB. Continuous integration, the automated build and test system that engineers shorten to CI, performs about 25 million code checkouts a day before anyone counts agent sandboxes and research jobs. Anthropic calculates that fetching all of that the naive way would generate more than 600 petabytes of Git traffic every day.

Git ships with partial fixes: shallow clones, sparse checkouts, internal mirrors, and snapshot copies kept in storage. Anthropic used all of them and, at the peak, operated several hundred internal mirrors. The mirrors bought time and added operating burden as the merge rate rose. Snapshots fell further behind the tip of the main branch between refreshes.

Scion moves the repository back to the server

Anthropic's replacement is called Scion, and it inverts Git's design. Clients assume the server is always reachable. When an agent commits or merges, the work happens on the server, so there is nothing to push.

The unit of work is a changeset, which lists the files that changed relative to the previous commit and points to their contents in blob storage. The cost of creating a changeset tracks the size of the change, however large the repository grows. Clients check out only the directories they request, and virtualized checkouts load a file when something actually reads it. Many jobs skip the checkout entirely and ask the server questions through an API, such as who last changed a line or how two commits relate.

Requests flow through three tiers: a daemon on each machine, regional caches, and authoritative servers backed by Google Spanner and cloud storage. Anthropic says each tier holds a cache hit rate above 99%.

Anthropic reports that the Scion footprint needs about one two-hundredth of the processor and memory capacity of its previous Git mirroring infrastructure. CI and research jobs check out code two orders of magnitude faster, in seconds instead of minutes, and CI alone freed 25% of its processor capacity.

The 200x baseline is Anthropic's own mirror fleet, so the number measures that setup and says little about what a typical Git host costs to run.

The merge queue is the quiet casualty

The most useful passage for a CIO sits in the lessons section. A merge queue lines up proposed changes and tests each one before it lands on the main branch, usually behind a global lock. Anthropic could not make that work at its volume. With hundreds or even thousands of changes merging while tests run, the company asks what continuous integration even means. It replaced the gate with automated testing after merge and before deployment.

Concurrent changes can now conflict. Anthropic says Claude typically fixes those conflicts faster than humans can identify them. That claim comes from the company that sells Claude.

Engineers became the people who say no

Anthropic reports that a team of seven built Scion and its code review tool, Antique, and migrated the monorepo in five months. Every engineer on the project now rates Claude better than themselves at writing code. Their job shifted to design tradeoffs and to telling Claude no.

The examples are specific. Several systems tried to find when a change merged by reading the commit log line by line when they could have asked the code review tool. An adversarial review pushed Claude to add excessive validation around a trusted internal service. Reviewers still examine every change request, but they judge the intent of each change and no longer read every line.

One data point carries the post's sales pitch. Engineers asked an early internal snapshot of Claude Mythos Preview to rewrite their hand-built ingestion pipeline, and Anthropic says the result used 20x less memory and ran 4x faster. The post doubles as a product demonstration, so read the numbers with that in mind.

What Anthropic left for the rest of us

Scion stays internal. Anthropic writes that it is by no means a suitable replacement for Git for general use, and that any adopter would need its own code review and code search to go with it. The company proposes a different path: Git platforms add comprehensive server-side calls, improve the negotiation protocol, and make sparse and shallow features work well together. It names GitHub, GitLab, and the open source community as partners.

Scion also has prior art. Anthropic credits Google's Piper and Meta's Sapling, two systems built for giant monorepos, as lessons it built on.

Anthropic also notes that its internal token spend exceeds what any other business would realistically choose to spend. My read: that caveat gives most companies more runway than the six-to-12-month forecast implies. My inference on the other side: the pressure moves to GitHub and GitLab next. I found no announced response from either vendor.