Anthropic Replaced Git With Its Own System After Daily Code Merges Hit 30,000

Anthropic has quietly done something few engineering organizations ever consider: it stopped using Git. In an engineering blog post published Monday, the company said it spent five months migrating its entire code monorepo to Scion, a version control system built in-house, after AI-generated code overwhelmed infrastructure that was designed for human contributors.

The numbers explain why. Daily merged code changes grew from roughly 2,500 in May 2026 to more than 30,000 by September — more than a tenfold increase in four months. The monorepo held nearly 100 GB of main-branch history with a compressed working copy above 25 GB, and continuous integration performed about 25 million checkouts per day, a figure that excludes developer environments, agent sandboxes and research jobs. Earlier reporting put Claude's share of Anthropic's merged production code above 80 percent.

Anthropic had already tried the standard escapes: partial downloads, repository snapshots and several hundred internal Git mirrors. Those bought time but kept failing — snapshots drifted behind the latest code, history-dependent operations got harder on partial repositories, and the mirror fleet stopped scaling once the repository outgrew individual servers' disks and memory. The bottleneck had moved: writing code was no longer the slow part, source control was.

Scion's design starts from a different assumption than Git's. Git presumes self-contained local repositories that can inspect history and check out revisions offline. Anthropic's clients — online agents and automated jobs — are always connected, so Scion assumes a reachable server. Servers create commits, perform merges and rebases, and answer history queries; each change is recorded as a server-created "changeset" listing files relative to its first parent. Clients keep a bounded cache and materialize working files on demand through sparse or virtualized checkouts.

The payoff, per the company's own measurements: Scion's CPU and memory requirements are one two-hundredth of the former Git-mirroring infrastructure, CI and research checkouts dropped from minutes to seconds — a 100-times improvement — and CI gained 25 percent more CPU capacity. Repository data is spread across independently scalable components, so no machine, including the servers themselves, holds a complete copy. The company also upgraded its code review system, known as Antique, as part of the migration.

Anthropic is explicit that Scion is not a general-purpose Git replacement and is only sharing the design approach, not the code. The trade-off is real: teams that depend on offline work lose Git's independence, and the system only makes sense when nearly every client is an agent with a network connection. The company also notes that latent demand for merging remains unsatisfied — clear any remaining bottleneck, it says, and output would immediately rise again.

The broader lesson lands on every engineering team now adopting coding agents. Version control, code review and CI tooling were all shaped by the assumption of human-scale commit volume. When a single lab's merge rate grows tenfold in four months, the infrastructure layer beneath software development becomes the next place AI economics get rewritten — and GitHub, GitLab and the rest of the forges now have a documented, public existence proof of what breaks first.