Anthropic replaced git with Scion after agent merges hit 30,000 a day — the checkout, not the commit, was the bottleneck

Anthropic replaced git with an in-house version-control system called Scion after Claude agents pushed daily merges in its monorepo from ~2,500 to over 30,000, and the company says source control had become its biggest development bottleneck. In an engineering post published October 13, 2026, a seven-person team described building Scion and migrating the entire monorepo off git in five months. CI checkouts dropped from minutes to seconds, and the infrastructure footprint shrank roughly 200x.

What shipped

The trigger was scale, not a git regression. Anthropic reports that its main-branch history alone is almost 100 GB, a compressed working copy is over 25 GB, and CI performs on the order of 25 million code checkouts per day — before counting developer environments, agent sandboxes, and research jobs. Fetched naively, that would be over 600 PB of git traffic per day, the post says. At peak the company ran several hundred internal git mirrors, plus shallow clones, sparse checkouts, promisor remotes, and blob-storage "seeds," and still could not keep up.

Scion inverts the client-server relationship that git assumes. The unit of work is a changeset — an entry in a write-ahead log that lists files changed relative to its first parent, plus content-addressed pointers to compressed file contents in blob storage. Changesets are created server-side and available everywhere instantly. There is no push step: every commit, rebase, or merge happens on the server, so there are no packfiles, no locking, and no cross-server coordination. The development lifecycle discourages refs (branches and tags) except for the main ref and release branches; moving one is a compare-and-swap on a single Spanner row.

The read path is where the design does its real work. It has three tiers — a scion-daemon on each compute node, a regional scion-cache, and the authoritative scion-server — all answering the same RPCs, with strict tiering so a failing regional cache cannot overwhelm the fleet. Each tier reports a cache hit rate above 99%.

There is no fully materialized copy of the repo anywhere, even on the server. Servers are essentially large caches that also compute derived data (file blame, path history, commit-graph skip lists) on the fly and write results back to Spanner. The client, scn, negotiates only the trees and blobs it actually wants, with resumable transfers and no single multi-gigabyte packfile response. Sparse and virtualized checkouts need no special support because the client never asks for trees it does not want.

Anthropic's reported results: checkouts for CI and research jobs dropped from minutes to seconds — "two orders of magnitude" faster — CI got roughly 25% of its CPU capacity back, and the infrastructure footprint shrank about 200x in CPU and memory versus the old git-mirror fleet. The same five months also produced Antique, a code-review surface with semantic approval constraints and a reworked merge queue, plus code search, per-directory IAM, and automated bisect/revert pipelines.

The recursion: Claude Mythos Preview itself rewrote the git-to-Scion ingestion pipeline, producing a version that used 20x less memory and ran 4x faster than the engineers' handwritten prototype. The post quotes the team's own assessment that "very often... the role of the engineers was to tell Claude 'no.'"

Why it matters

The takeaway is not "Anthropic killed git." At agent scale, the checkout — not the check-in — becomes the bottleneck, and Anthropic inverted the architecture to attack it. Every headline number (200x footprint, minutes-to-seconds checkouts, 25% CPU freed) flows from one decision: stop shipping a repository to every consumer and instead make source control a read-mostly, tiered-cache service where the unit of work is a server-side, content-addressed WAL entry and refs are a single database row.

That is a design pattern, not a product. If you build agent infrastructure, the merge rate is not the only thing that grows with agent adoption — the read load does, and it grows faster. Twenty-five million checkouts a day is a read-mostly workload; Anthropic made reads cheap object-level RPCs that hit a cache instead of materializing a repo. The prior art is explicit: the post credits Google's Piper and Meta's Sapling, and observers have pointed to Mononoke as the likely inspiration for the derived-data mechanism.

There is also a practical signal for anyone running a large monorepo with heavy CI: the post urges server-side git improvements — server-side diff/merge RPCs, smarter negotiation, better sparse and shallow composition — and says it hopes to work with GitHub and GitLab on them. Expect this class of problem to show up in more organizations over the next six to twelve months, per the post.

Caveats

Every number here is Anthropic's own account, so treat the 200x footprint, the 25% CPU reduction, and the >99% cache-hit figures as vendor-reported; no third party has independently validated them.

Scion is closed-source and explicitly not a general git replacement. The post states it is "by no means a suitable replacement for git for general use." It leans on Google Spanner and cloud blob storage, so it is not portable to other stacks, and offline support was dropped entirely. Per Scion engineer Nick G. on Hacker News (relayed by Simon Willison), "since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do." That trade-off is fine for Anthropic, but it is exactly why this is a case study rather than something you can adopt.

The post is also survivorship-biased: it says little about failed approaches or migration incidents, and the peak figure of 30,401 merges on September 9 comes from reading the post's chart rather than its prose.

Sources

• Redesigning source control at Anthropic (Anthropic engineering, Oct 13 2026)
• Anthropic reinvents source control as Claude floods the world with code (The Register)
• Simon Willison: Redesigning source control at Anthropic
• Anthropic ditches git for Scion (Frontier Now)
• XenoSpectrum: Anthropic Scion source-control redesign
• Hacker News discussion