Anthropic replaced Git with in-house Scion as daily merges hit 30,000

Anthropic has moved its entire monorepo off Git and onto Scion, a version control system built in-house by seven engineers over five months. The trigger, according to a dev.to write-up of Anthropic's engineering post, was scale: daily merges climbed from roughly 2,500 in May to more than 30,000 in September, and the Git setup could no longer keep up.

A monorepo that outgrew Git

The numbers behind the migration are unusual. Per the reporting, the main branch history of Anthropic's monorepo is close to 100 GB, and a compressed working copy exceeds 25 GB. CI alone performs around 25 million checkouts a day, before counting agent sandboxes and developer machines. Fetched the conventional way, that would translate into more than 600 PB of Git traffic daily.

Anthropic had been surviving on several hundred Git mirrors, but the workarounds degraded. Mirror seeds drifted behind the main branch between refreshes, and techniques like shallow clones and promisor remotes made history-dependent operations such as diff and blame fragile. As the dev.to author puts it, each patch introduced a new thing that needed babysitting.

How Scion works

Scion inverts Git's founding assumption. Git treats every client as offline and self-contained, with each clone holding the full repository. At Anthropic, nearly all clients are agents that are always online, so Scion moves the real work to the server: it creates commits, computes diffs, finds merge bases, and runs blame, while the client only lays files on disk and requests what it needs.

The unit of work is a "changeset" rather than a commit. A changeset records changes relative to its parent plus pointers to compressed file contents in blob storage. There are no packfiles and no pushing; commits and rebases happen server-side. Notably, the working copy is optional. A CI job that only needs a diff can ask the server and skip a checkout entirely, and when files on disk are unavoidable, the client performs sparse or virtualized checkouts, paying only for what it reads.

No fully materialized copy of the repository exists anywhere, not even on the server. Data is distributed and content-addressed, and because stored objects never change, they can be cached freely, letting the system scale horizontally.

The numbers, and their caveat

Anthropic reports that Scion shrunk its infrastructure footprint by a factor of 200 and cut CI checkouts from minutes to seconds. The dev.to author flags an important caveat: that 200x figure is measured against the previous setup of several hundred mirrors, which almost no other team runs. For teams merging a few hundred changes a day, Git remains well suited, and Scion is neither open source nor intended for general use.

Humans as the constraint layer

One of the more striking admissions in Anthropic's post, as highlighted in the write-up, is that engineers often found Claude writes better code than they do, and that their role frequently reduced to telling Claude "no." Agents would sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place. The dev.to author frames this as humans becoming the constraint layer rather than the code authors, a shift with uncomfortable implications for how engineering jobs are defined.

Why it matters

This is less a story about one company's tooling and more a warning about what AI-generated code volume does to standard infrastructure. Git was designed for humans writing code slowly, often offline. Agent-driven development removes both assumptions, and Anthropic reportedly expects other teams to hit similar walls within months. The company has also effectively handed the general problem to GitHub and GitLab, suggesting the industry's default version control stack may need rethinking. Equally significant is the redefinition of the engineer's job: reviewing, constraining, and vetoing machine output at a pace no human workflow was built for.