Anthropic has replaced git with a custom version control system called Scion, purpose-built for a world where AI agents produce code at a volume that breaks every assumption in traditional developer infrastructure. The numbers tell the story: from 2,500 merged code changes per day in May 2026 to over 30,000 per day by September, a 12x increase in four months with no plateau in sight. Git's core design — every clone a complete, self-contained repository — became a liability when the monorepo's history alone hit 100 GB and the company was performing 25 million code checkouts per day for CI alone. Scion inverts git's foundational assumption. Where git treats every client as a potential server capable of offline work, Scion assumes clients are always online and pushes all computation to the server. There are no packfiles, no local history, no full materialized copy of the repository anywhere. The unit of work is a "changeset" — a diff against a parent stored with content-addressed pointers to blob storage — which can be created with no locking, no cross-server coordination, and no ref movement. This makes commit creation scale with the size of the change, not the size of the repo. The architectural insight is that Anthropic was already using git as a centralized system with one authoritative copy, but paying the full cost of git's decentralized design at every checkout. Nearly all their clients — agents and CI jobs — are online by definition. Stripping away the offline assumption eliminated the 600+ petabytes per day of theoretical git traffic that naive fetches would have generated across their 25 million daily checkouts. Scion's sparse and virtualized checkouts mean clients only pay for what they actually read. Derived data structures — blame, path history, commit graph skip lists — are computed recursively from changesets and can be materialized eagerly or deferred. This lets Scion add new indices without changing the client, and the team reports adding new derived data types many times as the system grew. The entire stack runs on Spanner for metadata, with blob storage for file contents, and every component scales horizontally and independently. The five-month timeline from inception to full migration is itself a statement about AI-assisted development velocity. Anthropic built Scion, a code review surface (Antique), code search, a source control IAM system, and novel landing mechanisms — all in the same period. They explicitly credit Claude's code-writing capability as both the cause of the problem and the means of solving it. The team drew on prior art from Google's Piper and Meta's Sapling, both battle-tested monorepo systems, but designed specifically for the agent-first workload. The broader signal here is infrastructure divergence. Anthropic is the first company to publicly admit that standard developer tooling has hit a wall under AI-generated code volume, and they expect others to follow. Their merge rate growth curve shows no sign of flattening. If this pattern generalizes — and Anthropic claims it will — then git, CI systems, build tools, and code review workflows across the industry face a scaling cliff that current architectures cannot handle. What makes this genuinely generative rather than merely self-serving is the knowledge sharing. Publishing the design principles, data structures, and scaling decisions gives the industry a blueprint. But Scion itself is proprietary, and the competitive advantage of having solved this problem first — while competitors are still running git mirrors — is substantial. Anthropic is simultaneously building the future and gatekeeping the implementation.
