Anthropic Scion: Why Agents Outgrew Git at 30,000 Merges a Day

In May 2026, Anthropic merged about 2,500 code changes a day. By September it was merging more than 30,000, and Anthropic’s engineering team says git could no longer keep up, so it built its own source control system, Scion, and migrated its monorepo in five months. For engineering leaders whose coding agents are starting to produce more changes than people do, the reasoning behind that decision matters more than the system itself, which Anthropic says isn’t a general replacement for git.

How It Works

The numbers explain the break. The history of Anthropic’s main branch is almost 100 GB, a compressed working copy is over 25 GB, and CI performs on the order of 25 million checkouts a day. Fetched naively, Anthropic says, that would exceed 600 PB of git traffic daily. Shallow clones, sparse checkouts, and several hundred internal git mirrors bought time, but became harder to run as the merge rate climbed.

The insight was that git’s strengths didn’t apply. Git assumes every clone can work offline as a full repository. Nearly all of Anthropic’s clients are agents and jobs that are always online, so Scion is built on different principles, drawing on Google’s Piper and Meta’s Sapling:

• Server-side work. The server answers history and blame queries and creates commits, merges, and rebases. The client keeps a bounded cache and lays out files, so there’s nothing to push.
• Everything exposed as an API. CI jobs and tools can ask who changed a line or how two commits relate without a local repository, which makes automated bisect and revert pipelines practical.
• Cheap, lock-free commits. The unit of work is a “changeset” that records only changed files and pointers to content in blob storage. Cost scales with the size of the change, not the repo, and refs are discouraged except for main and release branches.
• Sparse checkouts by default. Clients request only the trees and blobs they need, through cache tiers that each hold a hit rate above 99%.

The results are large. Scion’s infrastructure is 200x smaller than the old git mirroring setup in CPU and memory, checkouts run in seconds rather than minutes, and CI alone gained 25% CPU capacity.

What This Means for You

A team of seven built Scion, plus a code review tool called Antique, and finished the migration while daily merge volume grew 10x. Anthropic says every engineer on it now rates Claude as better at writing code than they are, but the engineers’ job was often to tell Claude “no”, such as when systems tried to brute-force lookups that should have gone to the code review tool. Adversarial review helped, but Anthropic says it can be “myopic”, and engineers had to steer Claude away from over-engineering.

Other human-scale workflows are next. Anthropic dropped a global merge-queue lock in favor of post-merge and pre-deploy testing, letting concurrent changes conflict because Claude can usually fix them quickly. Agent-driven volume is already reshaping tooling elsewhere, as with Claude Code’s dynamic workflows running 1,000 agents. Simon Willison notes that Scion sounds like a reinvention of centralized source control along the lines of Subversion.

Source: Anthropic Engineering

Content created with AI assistance and reviewed for accuracy.