Anthropic builds Scion source control around always-online coding agents

Coding volume moves the bottleneck

Anthropic's engineering report describes an in-house source-control system called Scion. The company says its daily merged code changes grew from roughly 2,500 in May to more than 30,000 in September, increasing pressure on source control, builds and CI.

Anthropic reports migrating its monorepo from Git in five months. It also built related review, search and access-management surfaces. The account is about an unusually large, agent-heavy internal workload, rather than a general finding that ordinary repositories should replace Git.

Put repository operations on the service

Scion assumes clients are online. The server handles history queries, commit creation, merges and rebases, while clients maintain bounded local caches and materialize needed files. Source-control queries are exposed through APIs so some jobs can operate without a complete checkout.

Its changesets record changed files relative to a parent, with content-addressed file contents in blob storage. The report describes independently scalable components and sparse or virtualized working copies, designed to avoid repeatedly moving the whole repository.

Anthropic identifies coding agents and jobs as most of its clients. That requirement helps explain the design's dependence on a reachable service and its different tradeoffs from distributed offline repository work.

Diagnose the workload before changing the system

Franklin analysis: teams using coding agents should measure checkout, history-query and merge costs separately from model generation. A faster agent can increase infrastructure demand without making those operations faster.

For most teams, the first useful comparison is the existing system with narrower checkouts, caching and well-defined repository queries. If those measures stop meeting the workload, Scion's report supplies concrete design questions about which operations need local files and which can run through a service. Keep availability, audit history and recovery requirements beside throughput when considering that architectural tradeoff.