Anthropic has moved its internal monorepo off Git and onto Scion, a version control system it built from scratch because its AI coding agents were producing code faster than conventional tooling could absorb. In an engineering post published on October 13, 2026, the company said daily merges rose from about 2,500 in May to more than 30,000 by September, a tenfold jump driven largely by Claude-powered agents. The move is one of the clearest signs yet that agentic software development is starting to break infrastructure designed around human developers.
Why Git Stopped Working for an Agent Workforce
Git was designed for a world of human contributors who each keep a full, offline-capable copy of a repository. Anthropic argues that assumption no longer fits when almost all of its "clients" are agents and automated jobs that are always online.
The numbers in the post show the strain:
- The history of the main branch is close to 100 GB, and a compressed working copy exceeds 25 GB.
- The company performs roughly 25 million checkouts per day for continuous integration alone, not counting agent sandboxes and research jobs.
- Fetched naively, that volume would mean more than 600 PB of Git traffic every day.
- At peak, Anthropic was running several hundred Git mirrors, which became harder to operate as merge rates climbed.
The Register noted that GitHub has also linked recent availability problems to exponential growth in pull requests and commits, suggesting Anthropic is an early case of a wider pressure point.
How Scion Works
Scion is a centralized, server-heavy system that observers have compared to Apache Subversion more than to Git. The server handles history, blame, commits, merges and rebases, while clients keep only a bounded local cache and the files they actually need.
The basic unit of work is a changeset, stored as an entry in a write-ahead log. Rather than pointing to a complete tree, a changeset records which files changed relative to its parent, with content-addressed pointers to compressed data in blob storage. Branch references are updated through a compare-and-swap on a single Spanner row.
The system runs across three tiers: a daemon on each node, regional caches, and authoritative servers backed by Spanner and cloud storage. Anthropic says each tier maintains a cache hit rate above 99%. Crucially for agents, source control queries are exposed through an API, so many CI jobs and automated workflows never need a local checkout at all. The company says its automated bisect and revert pipelines depend on that access.
The reported results are substantial: Scion uses about 200 times less CPU and memory than the old Git mirror fleet, CI and research checkouts are around two orders of magnitude faster, and CI gained a 25% CPU capacity improvement.
Agents Built the Tool for Agents
Perhaps the most notable detail is who wrote the code. According to Anthropic, a team of seven engineers delivered Scion, a code review tool called Antique, code search, source control access management and several related systems, then finished the migration in about five months. Claude wrote most of the code.
An early Claude Mythos Preview snapshot rewrote the Git-to-Scion ingestion pipeline, cutting its memory use 20x and making it run four times faster. Claude agents were also assigned as dedicated monitors, each watching the health of a single component through telemetry and logs.
The human role shifted toward architecture, design trade-offs and restraint. Anthropic wrote that engineers often had to "tell Claude 'no'" when agents broke core system constraints or over-engineered solutions in response to adversarial reviews. Reviewers now concentrate on the intent of a change rather than every line.
Antique supports semantic approval rules that can distinguish comment-only edits from consequential logic changes. There is no global merge-queue lock; instead, Anthropic relies on post-merge and pre-deploy automated testing, with Claude typically resolving conflicts from concurrent merges.
Why It Matters
Most discussion of agentic AI focuses on what agents can do. Scion is a reminder that the harder problem may be everything around them. When autonomous coding agents raise output by an order of magnitude, the bottleneck moves from writing code to storing, reviewing, testing and safely landing it.
Several implications stand out for engineering leaders:
- Developer tooling will be redesigned for machine clients. API-first access, sparse checkouts and server-side operations matter more when agents, not people, are the main users.
- Review is becoming policy-driven. Semantic approval rules and automated testing are replacing line-by-line human inspection at high volume.
- Human judgement moves up the stack. Anthropic's experience suggests engineers become architects and editors who constrain agent behaviour rather than primary authors.
Large technology firms have built bespoke systems before, including Google's Piper and Meta's Sapling, but those were responses to huge human engineering organisations. Anthropic reached a comparable breaking point with a far smaller team, because its agents supplied the volume.
What Comes Next
Scion will not be open-sourced. Anthropic says the system is tailored to its environment and is not a general Git replacement. Instead, it published the design so others can apply the ideas, and said it intends to work with GitHub, GitLab and the open-source community on the underlying problems. The company argues that similar scaling issues will soon reach other organisations adopting coding agents at scale, which makes how established Git platforms respond the next thing to watch.
