Anthropic Moved Its Monorepo Off Git to Scion as Daily Merges Grew From 2,500 to 30,000

Anthropic has moved its monorepo off git onto an in-house system called Scion, and the reason it gives is arithmetic: merges to the repository grew from about 2,500 a day in May 2026 to more than 30,000 a day by September. The post describing the move, published October 13, is useful less for the architecture than for what it admits about the limits of a system built for agents.

What the post says was built

"Redesigning source control at Anthropic" is credited to Nick G. and the source control team, which the post puts at seven people. It says the bottleneck is "all of the developer infrastructure built for human scale: source control, builds, and CI." The team migrated the monorepo from git to Scion in five months, and wrote a companion code review tool, Antique.

The design is a write-ahead log of changesets, each recording changed files relative to its first parent instead of a pointer to a whole tree. Indices for blame, path history and the commit graph are derived data. A client, scn, and a daemon, scion-daemon, keep a bounded object cache on each node. Regional scion-cache instances sit in front of authoritative scion-server instances, which read from Spanner or cloud storage.

All the figures below are Anthropic's own. The post gives no third-party measurement.

Measure	Figure in the post
Main-branch history	almost 100 GB
Compressed working copy	over 25 GB
CI code checkouts per day	about 25 million
Git traffic if fetched naively	over 600 PB a day
Cache hit rate at each tier	over 99%
CPU and memory vs. old git mirrors	about 200x smaller

1. May 2026 (about)2500 merges
2. September 2026 (over)30K merges

Source: Anthropic Engineering, Redesigning source control at Anthropic, accessed 2026-10-13

Why git was the problem

Git gives every clone a full copy of the history, which suits people who work offline. Anthropic's clients are mostly agents and CI jobs that are always online. Without a change, the post says, the 25 million daily checkouts would mean over 600 PB of git traffic a day, and the team had run several hundred internal git mirrors at peak to cope.

The post credits Google's Piper and Meta's Sapling as prior art, which is the honest description: sparse, virtualized checkouts and a server-side history are old ideas at the largest companies. What is new is the reason, which is volume produced by agents. Two studies of coding agents covered here found the same pressure from the other side, with commits rising far faster than releases.

What Anthropic concedes

The sentence only a reader of the post would find is the one about the merge queue: Anthropic could not replicate a global repository lock, so concurrent changes can conflict, and the team leans on post-merge and pre-deploy testing instead. The post also says refs are discouraged except for main and release branches, and that changesets record only first-parent changes, which works because of trunk-based development.

It is blunt on scope: "Scion is by no means a suitable replacement for git for general use." It adds that "it's difficult to offer Scion in a form that would be broadly useful to other organizations", so no release is promised. Anthropic commits instead to partnering with GitHub, GitLab and the open-source community.

One line is easy to skim past. The post says Claude Mythos Preview rewrote the ingestion path, using 20x less memory and running 4x faster than the handwritten prototype. That is a vendor figure about the vendor's own model, with no code published to test it against.

The case against reading it as a trend

The Hacker News thread split. Commenter preommr warned that "everybody FOMOs into thinking they need to do the same." Commenter amluto argued that 30,000 merges a day is "not actually very many per employee" and suspected the real pain was that git clone is slow. Commenter bix6 questioned the 600 PB figure, which is a hypothetical about naive fetching, not a measured load.

Commenter amluto also objected to server-side commits, because building history locally is a benefit they do not want to lose. Another commenter, steveklabnik, suggested a git bridge so teams need not "swap services and move to something non-git at once."

The review question was raised directly. Commenter jareds asked whether Anthropic ships code without human review. The post's own answer is that "engineers still review the intent of every change", while agents "sometimes violate system constraints or over-engineer."

The Torvalds patches written with Claude Code and Scion sit at opposite ends of the same question: what does infrastructure look like when most changes are not typed by a person?

What to watch

The post also says Anthropic's internal token spending is far above what most businesses would spend, which makes its setup unusual by the company's own account. The number to watch is whether any outside host ships the sparse, first-parent design for git itself, since Anthropic has said that is where it will bring the ideas.