Anthropic Scion: Why Claude Needed New Source Control

Anthropic's new source control system, Scion, is a fascinating look at what happens to software engineering when AI writes most of the code. In an engineering post published October 13, Anthropic explained that daily code merges to its main repository grew from about 2,500 in May 2026 to more than 30,000 in September. Git, the tool almost every developer uses, could not keep up at that volume, so a small team built a replacement, with Claude writing most of the code.

• Scale: merges rose from around 2,500 per day in May to over 30,000 per day in September 2026, according to Anthropic.
• Efficiency: Scion uses roughly 200x less CPU and memory than the git mirror infrastructure it replaced.
• Speed: CI and research checkouts dropped from minutes to seconds, and CI gained about 25% CPU capacity.
• Team: seven people built Scion and its companion tools, and migrated the monorepo off git in five months.

Why Did Anthropic Need a New Source Control System?

The numbers explain it. Anthropic's main branch history is almost 100GB, and a compressed working copy is more than 25GB. Its continuous integration system alone performs around 25 million checkouts per day, before counting developer machines, agent sandboxes and research jobs. Fetched naively with git, Anthropic says that would add up to more than 600 petabytes of traffic every day.

Git's decentralized design, where every copy of a repository carries the full history, is a brilliant fit for most projects. As The Register points out, though, the largest engineering organizations have long moved toward centralized systems for giant monorepos: Google runs Piper, Meta built Sapling, and Microsoft created the Virtual File System for Git. Anthropic arrived at the same place, just much faster, because AI coding agents multiplied its code output in a matter of months.

How Does Scion Work?

Scion is server-authoritative. Instead of each client holding the full history, a central service handles history, blame, commits, merges and rebases. According to Anthropic's post, the design has a few key pieces:

• Changesets: each unit of work stores the files it changed plus content-addressed pointers to the underlying data.
• On-demand checkouts: clients fetch only the files they actually read, using sparse and virtualized checkouts.
• Three cache tiers: a per-machine daemon, regional caches and authoritative servers, each with a cache hit rate above 99%.
• Simple, strong storage: references are updated with a compare-and-swap on a single database row, with file contents in cloud storage.

That architecture is why a checkout that once took minutes now takes seconds. When an agent only needs a handful of files, it no longer has to drag 25GB across the network first.

What Role Did Claude Play in Building Scion?

This is the most striking part of the story. Anthropic says Claude wrote most of Scion's code, along with a new code review tool called Antique, a code search service and a fine-grained access control system. In one example, an early internal snapshot of Claude Mythos Preview rewrote the pipeline that imports git history into Scion without further instructions. The rewrite used 20x less memory and ran 4x faster while keeping its ordering guarantees intact.

The human engineers' work shifted toward design: setting constraints, reviewing the intent behind changes and rejecting approaches that broke the system's rules. Anthropic even runs dedicated Claude instances that watch individual components through telemetry and logs, alongside traditional alerts. It is a concrete example of the agentic coding workflow scaling up from individual developers to core infrastructure.

Will Scion Be Open Source?

Not as a product. Anthropic says Scion diverges too far from git, and depends on its own code review, search and CI systems, to offer as a general replacement. Instead, the company plans to share its design thinking so others can apply the ideas to git, and says it hopes to work with GitHub, GitLab and the open-source community on that.

The Bigger Lesson for Developers

Most teams are nowhere near 30,000 merges a day. But the trend line matters: as coding agents get faster, the bottleneck moves from writing code to everything that happens after, including review, testing, merging and deployment. Anthropic's takeaway is that small teams can now finish large infrastructure projects quickly, and that human-scale habits, like a single global merge lock, may need rethinking for agent-scale development. For more on where AI development tools are heading, browse our AI coverage.

Sources: Anthropic Engineering, Redesigning source control — October 13, 2026; The Register, Anthropic reinvents source control — October 13, 2026.