In-depth analysis · About 5 min read
Why Anthropic built Scion for an agent-scale monorepo
The internal source-control system moves computation to servers and retrieves immutable objects on demand; the engineering account is a workload-specific case study, not a general replacement for Git.
- Source published
- Published here
- Source checked
Anthropic’s October 13 engineering account describes migrating its monorepo to Scion, an internal source-control system, in five months. The article was updated October 15 to credit earlier work. Its central argument concerns online agents and unusually frequent checkouts, rather than a claim that ordinary developers should abandon Git.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Anthropic reports daily merged changes growing from about 2,500 in May to more than 30,000 in September; seven engineers built Scion, Antique and related systems during the migration.
- Scion stores incremental changesets and content-addressed objects, while servers handle history, merges and derived indexes; clients fetch only needed data.
- Reported cache hit rates exceed 99% at each tier. Resource and speed gains are Anthropic’s observations on its infrastructure, not AISubHub benchmarks.
The workload that made copying a repository expensive
The company describes almost 100 GB of main-branch history, a compressed working copy exceeding 25 GB, and roughly 25 million daily CI checkouts. Mirrors, seeds and partial-clone techniques helped, but maintaining fresh snapshots and usable history became operational work of its own. These figures explain the design pressure: repeatedly moving repository state can consume resources even before an agent edits a file. Most clients were already online jobs using one authoritative repository. Our analysis is that the mismatch lies between that centralized, always-connected workload and the costs of equipping every client for independent offline operations. It does not invalidate Git’s decentralized design for ordinary repositories.
Changesets move work to the service
Scion puts commit creation, rebases, merges, history queries and blame on servers. A client keeps a bounded object cache and materializes requested files. A changeset records files changed relative to its first parent and references compressed content-addressed blobs; it is not a complete local repository. Additional indexes supply path history, blame and commit-graph traversal. The article calls these derived data, computed recursively from parent changesets either eagerly or when needed. This separation lets engineers change an index without changing clients. Our reading is that the useful lesson is separating durable change records from query-specific acceleration, rather than copying a particular database choice into every development stack.
Immutable objects enable a tiered checkout path
Clients request missing trees and blobs instead of negotiating one large packfile. On a machine, scion-daemon shares objects between jobs; a regional scion-cache serves those objects, with authoritative scion-servers behind it using Spanner or cloud storage. Because an object’s content does not change, caches can answer without coordinating for every read. Sparse or virtualized checkouts simply avoid requesting unnecessary objects. Anthropic reports more than 99% cache hits at each tier, an infrastructure footprint about 200× smaller in CPU and memory requirements than its previous Git mirrors, and a 25% CI CPU-capacity gain. These are deployment-specific observations; they are neither universal speed guarantees nor results of our own testing.
Human decisions and safety controls remain part of the system
The account says engineers still review every change request, emphasizing intent and system constraints. It gives examples where agents pursued expensive history scans or added unnecessary defensive complexity. The system also supports directory-level access controls and semantic review requirements. For very high merge concurrency, the team invested in post-merge and pre-deployment tests instead of reproducing a global merge-queue lock, accepting that concurrent changes can conflict. Our interpretation is that faster code production increases the importance of choosing invariants, failure boundaries and review policies. Automated implementation does not itself demonstrate that an authorization boundary or a migration is safe.
What another team can actually take away
Anthropic explicitly says Scion is unsuitable as a general Git replacement and difficult to distribute as a useful standalone product. Its bespoke review, search and CI components, extraordinary monorepo scale and internal token budget are part of the result. The transferable ideas are narrower: expose useful repository queries through services, avoid redundant full checkouts, and measure where data movement dominates. The company proposes sharing techniques with existing Git platforms rather than asking the industry to migrate wholesale. For a smaller team, our practical conclusion is to measure checkout traffic and CI bottlenecks before changing source control; an internal case study is evidence to investigate, not a product recommendation or a promised saving.
Why it matters
The article connects agent productivity with infrastructure costs that model benchmarks miss. It offers concrete mechanisms to evaluate when repository access, history queries or CI setup become more expensive than the code-writing step itself.
Limits and availability
All operating numbers come from Anthropic’s own account. We did not deploy Scion or reproduce its measurements. The article announces an internal migration, not a generally available service, open-source release or new Claude subscription entitlement.
Sources and evidence
Source published: Oct 13, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.