Anthropic published Redesigning source control at Anthropic, describing Scion, the in-house version control system it migrated to in five months. The trigger was volume: in May 2026 Anthropic averaged about 2,500 code changes merged per day, and as of September it was over 30,000, a 12x increase. Source control became its biggest development bottleneck: coding is no longer the constraint, Anthropic says, developer infrastructure built for human scale is. It expects growth to continue, since latent demand for merging code is "practically infinite".
Why git stopped fitting
Git is decentralized: every clone is complete and works offline. Anthropic's monorepo has a main-branch history of almost 100 GB and a compressed working copy of over 25 GB. CI alone runs about 25 million checkouts a day, before developer environments, agent sandboxes and research jobs, over 600 PB of naive git traffic per day.
Shallow clones, sparse checkouts, promisor remotes and several hundred internal mirrors at peak kept things running, but grew harder to operate as merges accelerated. Anthropic's insight: it used git as a centralized system, yet nearly all its clients are agents and jobs that are online by definition.
Scion's design
Scion rests on six principles: clients are always online and fetch data on demand; the server answers history queries and creates commits, merges and rebases, while the client keeps a bounded object cache; every query is exposed through an API, so jobs such as automated bisect and revert often need no local repository; everything scales horizontally, with no fully materialized copy and no packfiles; commits are cheap, incremental and lock-free; and the working copy is optional and sparse by design.
The unit of work is a changeset, an entry in a write-ahead log that lists changed files and points to content-addressed, compressed contents in blob storage. Moving a ref is a compare-and-swap on a single Spanner row. Derived data such as blame, path history and search indices comes from one recursive mechanism invisible to the client.
The scn client asks for the trees and blobs it wants, and each request is for an immutable object any cache can answer. Requests go to scion-daemon on the node, then to regional scion-cache servers, then to authoritative scion-servers reading Spanner or cloud storage. All tiers answer identical RPCs, and tiering is strictly enforced to isolate failure domains.
Results
Each tier has a cache hit rate above 99% and no long-lived state. Scion's footprint, in CPU and memory, is 200x smaller than the previous git mirroring infrastructure. CI and research jobs check out code two orders of magnitude faster, in seconds rather than minutes, and CI alone gained 25% in CPU capacity.
What Anthropic learned
Work started around the time Claude Mythos Preview and Claude Tag became available internally. An early internal Claude Mythos Preview snapshot rewrote the hand-written git-to-Scion ingestion prototype with no further instructions: 20x less memory, 4x faster, with implicit invariants preserved. A team of seven built Scion, the Antique code review surface and half a dozen other systems and fully migrated in five months while merges per day rose tenfold; this previously took years and many more engineers.
All the engineers now rate Claude better than themselves at writing code, but they supplied the design and trade-offs, often by telling Claude no. Agents brute-force read the commit log to find when a change merged instead of asking the code review tool, and adversarial review sometimes over-engineered validation. Team members still review every change request, examining intent rather than every line. Quality rests on hermetic simulations, fuzzing and rate limiting, and Claude found optimizations such as an ancestry-query algorithm that missed an index.
Repositories carry per-directory permissions, and code review expresses semantic constraints on approvals, such as comment-only changes versus sensitive logic. For operations, Claude Tag runs "superfans": Slack channels where Claude monitors one component, such as scion-daemon or the Antique merge queue, catching subtle issues beyond traditional alerts.
The merge queue was rethought too. When hundreds or thousands of changes merge while tests run, a conventional queue stops making sense, and the industry-standard global repository lock could not be replicated. Anthropic invested in post-merge and pre-deploy testing instead. Concurrent changes can conflict, but Claude typically fixes conflicts faster than humans can identify them.
Not a git replacement
Anthropic says Scion is not a replacement for git for general use. It makes sense only for enormous monorepos with tens of thousands of engineers or swarms of agents, and Anthropic's token spend far exceeds what other businesses would want. It is also hard to offer broadly, since it diverges from git and needs its own code review, search and CI.
Anthropic expects this class of problem to become common within 6 to 12 months and suggests git platforms adopt comprehensive server-side RPCs (perhaps standardized across forges), a different negotiation protocol, and sparse and shallow features that compose well. It thanks the git community for nearly two decades of work and wants to partner with GitHub, GitLab and the open-source community. It appeared the same day as Anthropic's distillation defenses.