30,000 merges a day! Anthropic built Scion to replace git, and migrated in 5 months

Hi everyone, it's Shii-chan! Anthropic built its own version control system!

Anthropic Engineering anthropic.com

What was announced?

This is the Anthropic Engineering post "Redesigning source control at Anthropic". Merge volume grew from about 2,500 code changes per day in May 2026 to over 30,000 per day by September, and source control became the main development bottleneck.

So Anthropic built Scion, an in-house version control system, and migrated its monorepo from git in five months. It also built Antique (a code review tool), code search, and a source control IAM system. The post says a seven-person team did this work.

Why it matters

When most clients are agents and jobs rather than people, git's assumptions stop fitting. The post lists these scale problems:

• The main branch history is almost 100 GB, and a compressed working copy is over 25 GB
• There are about 25 million CI checkouts per day. Fetched naively, that would be over 600 PB of git traffic daily
• At peak there were several hundred internal git mirrors, which got harder to maintain as merge rates rose, and seeds drifted from the tip of main

What changes

The reported results:

• Each cache tier maintains a hit rate above 99%
• CPU and memory footprint is about 200x smaller than the previous git mirroring infrastructure
• CI and research checkouts take seconds instead of minutes, about two orders of magnitude faster
• CI alone gained a 25% CPU capacity win

Dive Deep

There are six design principles:

• Clients are always online: most are agents and jobs, so the design assumes a reachable server and fetches data on demand
• Client-server asymmetry: the server handles history, blame, commits, merges, and rebases, while the client keeps a bounded local object cache and manages the working copy
• Source control as an API: tools such as automated bisect and revert need no local repository
• Horizontal scalability: no server holds a full copy of the repo, and no packfile maintenance is required
• Cheap, lock-free commits: commits happen on the server, so there is nothing to push. Refs are discouraged except for main and release branches
• Optional, sparse working copies: CI can request a diff or write changes via API without checking out code

Key architecture points:

• Changesets: the unit of work, stored in a write-ahead log. Each lists the files changed relative to its first parent, with content-addressed pointers to compressed blobs. Moving a ref is a compare-and-swap on a single Spanner row
• Derived data: blame, path history, and skip lists for the commit graph are computed recursively from parent values through one generic mechanism, eagerly or deferred
• Fetch protocol: unlike git, negotiation covers only the trees and blobs the client wants. The client walks from the root tree, and transfers are resumable without a large packfile
• Three tiers: scion-daemon per node, scion-cache per region, and the authoritative scion-server reading from Spanner or cloud storage. Tiering is strictly enforced to isolate failures

Lessons involving Claude:

• An early Claude Mythos Preview snapshot rewrote the hand-built git-to-Scion ingestion pipeline with no extra instructions. The new version used 20x less memory, ran 4x faster, and kept the existing ordering and data-structure invariants
• Engineers focused on architecture and design tradeoffs, and often had to reject patterns Claude produced, such as linearly scanning commit logs
• Testing and security: hermetic simulations, fuzzing, and rate limiting, plus directory-level permissions and semantic approval rules
• Operations: Claude "superfan" agents each monitor one component, such as scion-daemon, scion-cache, or the Antique merge queue
• Merge queues: a global-lock merge queue doesn't suit agent scale. Anthropic relies on post-merge and pre-deploy automated testing, and Claude typically fixes conflicts faster than humans can find them

Anthropic says Scion is not a general git replacement. It requires migrating the whole development lifecycle, and its tools are highly bespoke. It shares the ideas so git platforms and clients can adopt parts of the model, and plans to work with GitHub, GitLab, and the open-source community.

Wrap-up

• Merge volume jumped from about 2,500 per day in May to over 30,000 per day by September
• Anthropic built the server-centric, sparse Scion in place of git and migrated in five months
• Cache hit rates are above 99% and resource use is about 200x smaller
• It's not a general git replacement; the goal is to share ideas

This one is for anyone interested in infrastructure design for an era when agents write lots of code!