Anthropic Replaced Git with Scion: Source Control Built for AI-Agent Volume

What happened

Anthropic published an engineering write-up on why it replaced Git for its monorepo. Simon Willison highlighted it and quoted the headline figure: around 2,500 code changes merged per day in May 2026, and over 30,000 per day by September.

Willison guesses that the newest model generation is a key driver of the jump. He is explicit that his chart marks public release dates, not when models became available internally. Treat the timing as suggestive, not proven. Anthropic's own account is that change volume rose more than tenfold in a few months, and the tooling had to follow.

Why Git stopped fitting

Git gives every clone the full repository, which suits offline, distributed human work. Anthropic's clients are mostly agents and jobs that are always online, so that design became a cost. According to the write-up, the main branch history is nearly 100 GB, a compressed working copy is over 25 GB, and CI performs about 25 million checkouts a day. Fetched naively, that would be over 600 PB of Git traffic daily.

Shallow clones, sparse checkouts and seeding helped. At peak, though, Anthropic ran several hundred internal Git mirrors, and operating them got harder as merge volume grew.

What Scion changes

• Always-online clients: the server handles history, blame, commits, merges and rebases. The client keeps a bounded local cache.
• Changesets in a write-ahead log: each records only the files changed relative to its first parent, with content-addressed pointers to blobs, and is created server-side.
• Derived data: blame, path history and commit-graph skip lists are computed from changesets, so new indexes need no client change.
• Minimal refs: discouraged except for main and release branches. Moving a ref is a compare-and-swap on a single Spanner row.
• Optional, sparse working copies: many CI and write tasks run through the API with no checkout at all.

The architecture is tiered: the scn client talks to a per-node scion-daemon, which fetches from regional scion-cache servers, backed by an authoritative scion-server. Anthropic reports cache hit rates above 99% per tier, and checkouts for CI and research dropping from minutes to seconds.

Review is the other half

Anthropic also built a code review tool, Antique, with a merge queue. It can express semantic approval rules, such as telling comment-only changes apart from changes to sensitive logic. The global merge-queue lock was replaced with post-merge and pre-deploy automated testing.

The write-up says engineers' main role shifted toward setting intent and rejecting designs that violate system constraints, rather than reading every line. It also notes that Claude-driven adversarial review can over-engineer in response to hypothetical concerns, so engineers had to steer it back.

Why it matters to builders

Most teams are nowhere near this scale, and nobody needs to leave Git. The useful lesson is where the pressure lands once AI-assisted work speeds up. Anthropic's bottlenecks were checkout cost, mirror operations, review throughput and merge serialization. Typing code was not among them. If you adopt agentic workflows, expect strain on your CI fan-out, your review policy and your merge process before you hit any model limit.

One detail is worth noting. Anthropic reports that Claude rewrote a hand-built Git-to-Scion ingestion prototype and produced a version using 20x less memory and running 4x faster. A team of seven built Scion, Antique and about six related systems, and the full migration took five months.