Anthropic Scion: version control built for agents

Anthropic has retired Git inside its monorepo in favor of Scion, a server-centric version control system it wrote itself, where servers own history, blame and merges and clients keep nothing more than a bounded local cache. The forcing function is agent output: in the company's engineering write-up, merged changes went from roughly 2,500 per day in May 2026 to more than 30,000 per day in September 2026. Seven people finished the migration in five months, while the volume they were migrating grew tenfold underneath them.

The wall Git ran into

The limits Anthropic names are physical rather than aesthetic. History on the main branch alone is close to 100 GB, and a compressed working copy ran past 25 GB. CI generated on the order of 25 million checkouts per day. At peak the company ran several hundred internal Git mirrors, which got harder to operate as the merge rate rose.

The assumptions Scion inverts

Scion starts from premises Git does not share. Clients are assumed to be always online, because agents and jobs are online by definition. Servers answer every query through an API, and operations are meant to scale with the size of a change rather than the size of the repository. No fully materialized copy of the repo exists anywhere in the system.

Three layers sit on top of each other: scion-daemon for node-level storage, regional scion-cache servers that stream objects out of memory, and authoritative scion-servers reading from Spanner or cloud storage. A changeset stores no tree pointers; it holds the list of files that changed relative to its first parent plus content-addressed pointers into compressed blob storage. Commits are lock-free, settled by a compare-and-swap on a single Spanner row. Checkouts come sparse, covering only the directories a client asks for, or virtualized, materializing files on demand.

The gains Anthropic publishes

Every number below is Anthropic's own internal measurement, with no published methodology and no independent check. The company puts the infrastructure footprint for source distribution at 200x smaller in CPU and memory than its earlier Git mirroring. CI checkouts are described as two orders of magnitude faster, seconds instead of minutes, which freed up a 25% CPU capacity win in CI. Each cache tier is said to hold a hit rate above 99%.

Claude shows up as a tool on its own migration. A Claude model that the post calls Mythos Preview rewrote the handwritten Git-to-Scion ingestion pipeline, producing a version that used 20x less memory and ran 4x faster while preserving the implicit ordering and data-structure invariants.

Designed for agents, not for review queues

The reasoning matters more than the plumbing. Workflows built for human pace no longer fit, Anthropic argues, and a proper merge queue in particular makes little sense at agent scale. Concurrent changes can still conflict, but the post claims Claude typically fixes those conflicts faster than people can even spot them. Refs such as branches and tags are deliberately discouraged, apart from the main ref and release branches.

Paired with Scion is Antique, the code review tool the same seven-person team shipped. It carries semantic constraints on approvals, separating comment-only edits from sensitive logic changes. The post stops there and hints at a separate article on Antique.

Not a product, and not offline

Scion is “by no means a suitable replacement for git for general use.”

That is Anthropic's own verdict on it. The system presumes bespoke review, search and CI tooling, and the company calls it difficult to offer in a form broadly useful to other organizations. Rather than a product, it promises to share its thinking and to work with GitHub, GitLab and the open source community, pointing to comprehensive server-side RPCs as one idea Git platforms could adopt. Google's Piper and Meta's Sapling are credited as prior art.

Offline use is not addressed in Anthropic's text at all. Simon Willison, writing on October 13, 2026, notes the absence of offline support and quotes author Nick G. explaining that Claude is by far the heaviest user of Scion and is necessarily online. Willison's further reads — that newer model generations lifted coding velocity, and that the derived-data idea echoes Mononoke — are his commentary, not claims in the Anthropic post.

What is not verified

The cost of the rebuild is not disclosed, and neither is any measure of code quality at more than 30,000 merges per day. The post does not break out how much of that volume comes from agents versus humans. It says only that the company expects no slowdown, because every bottleneck removed leads straight to more code being produced.