Anthropic Replaces Git with Scion to Support 30,000 Daily Agent Merges
Anthropic migrated its monorepo from Git to Scion, a custom version control system built for always-online coding agents. The architecture uses virtualized checkouts and server-side log operations to handle over 30,000 merges and 25 million CI checkouts per day.
Why it matters
Understand the architecture patterns needed to prevent Git and CI bottlenecks when scaling autonomous Claude Code worktrees across engineering teams.
TL;DR
- 01Git's offline-first architecture creates massive I/O and network overhead when scaled to tens of thousands of automated agent merges.
- 02Scion replaces git pushes with server-side write-ahead log changesets and on-demand file materialization.
- 03Programmatic source control APIs allow CI pipelines to run diff, blame, and automated reverts without downloading full repository trees.
Key facts
- Daily Merges in May 2026
- 2,500
- Daily Merges in September 2026
- Over 30,000
- Main Branch History Size
- Almost 100 GB
- Compressed Working Copy Size
- Over 25 GB
- Daily CI Checkouts
- ~25 million
- Migration Timeframe
- 5 months
The Breakdown of Distributed Git at Agent Scale
When Claude autonomously iterates on loops, micro-optimizes codebases, and spins up background services, the primary bottleneck becomes human-scale developer infrastructure. Anthropic's monorepo main branch reached 100 GB in history, with a 25 GB compressed working tree. Handling 25 million daily checkouts across CI, sandboxes, and research jobs threatened to generate over 600 PB of Git network traffic per day. Running several hundred internal Git mirrors failed as monorepo snapshots diverged faster than refresh cycles could keep up.
Scion Core Architecture Principles
Anthropic established five core principles for agentic source control:
- Client-Server Asymmetry: Clients do not compute diffs or blames locally. The server executes merges, rebases, and commit creation; the client only manages a bounded disk cache.
- Server-Side Changesets: The unit of work is a changeset written directly to a write-ahead log. There is no
git pushstep; changes land immediately on the server with lock-free semantics. - Virtualized and Sparse Checkouts: CI pipelines and agents query repository states via HTTP APIs. When disk access is strictly required, files are materialized on demand, meaning jobs only pay bandwidth for files they read.
- Ref Minimization: Long-lived branches and global refs are discouraged to eliminate distributed state agreement overhead across server clusters.
Implications for Agentic Engineering Teams
While Anthropic built Scion alongside Antique (code review) and custom code search for internal use, the architectural pattern highlights an impending industry shift: standard Git workflows degrade quickly once multi-agent orchestrations submit continuous pull requests.
✓ When to use
- Autonomous multi-agent loops generating high-velocity changes that saturate Git locks and storage mirrors.
- Large-scale monorepo CI infrastructures executing millions of daily sparse test runs.
✕ When NOT to use
- Standard development teams where merge volumes remain under a few hundred PRs per day.
- Projects requiring full offline repository manipulation and distributed peer-to-peer syncing.
What to do today
- Audit CI checkout depths: enable shallow clones (`git clone --depth=1`) and sparse checkouts in agent build jobs to cut disk footprint.
- Profile multi-agent worktree operations to identify local Git mirror locks and cache saturation before scaling parallel agent loops.
- Design agent pipelines to query git metadata via hosted platform APIs rather than executing full repository clones.
What the community says
“Practically speaking, it’s difficult to offer Scion in a form that would be broadly useful to other organizations. It diverges significantly from git, meaning companies would need to migrate...”
“Honest question: at this point, what do people do at Anthropic, if this post is to be believed? Sounds like Claude loops on itself and must be pushing millions of lines of code changes per day.”