Anthropic Replaced Git With Scion After Merges Hit 30,401 in a Day, and Says You Should Not Copy It

Anthropic merged about 2,500 code changes a day in May 2026 and more than 30,000 a day by September, and the company says Git could not keep up. Its answer is Scion, an internal source-control system built by a team of seven and described in an Anthropic engineering post published October 13 and updated October 15.

The task the post describes is a plain one: let tens of thousands of agent jobs read and change one repository without waiting on a clone. Everything below is Anthropic's own account. No outside party has measured Scion.

What broke at 30,401 merges a day

Simon Willison read the chart in the post and put the peak at 30,401 merges on September 9, 2026. Anthropic says main-branch history is almost 100 GB and a compressed working copy is over 25 GB.

Continuous integration alone pulls about 25 million checkouts a day, before developer environments, agent sandboxes and research jobs. Fetched naively, the post estimates that would be over 600 PB of Git traffic daily.

1. May 2026 (average)2500 merges per day
2. September 2026 (peak day, Sept. 9)30K merges per day

Source: Anthropic Engineering, Redesigning source control at Anthropic, accessed 2026-10-16

The chart compares an average with a single peak day, so the gap is wider than the post's own "over 30,000 per day" September figure. Even at that figure the rise is more than tenfold.

The workarounds before Scion were shallow clones, sparse checkouts, promisor remotes, internal mirrors and seeds in blob storage. At peak Anthropic ran several hundred internal Git mirrors, and the seeds drifted from main between refreshes.

Why Git was the wrong shape

The post's diagnosis is that Git assumes offline clients holding full copies. Anthropic used it as a central system where nearly every client was an always-online agent or job.

Scion therefore starts from the opposite assumption. The client fetches on demand, keeps a bounded object cache and leaves history, blame, merge and rebase to the server. Commits, rebases and merges happen on the server, so there is nothing to push.

The unit of work is a changeset, an entry in a write-ahead log that lists files changed against its first parent. Moving a ref is a compare-and-swap on a single Spanner row, kept separate from creating the changeset. Blame and path history are "derived data", computed through one generic mechanism rather than stored in the commit.

Layer	Role in Scion	Source claim
scion-daemon	One per node, serves the whole node	Starts from a refreshed snapshot
scion-cache	Regional, streams from memory	Over 99% hit rate per tier
scion-server	Authoritative, reads Spanner or cloud storage	No full repo copy anywhere

The numbers Anthropic reports for itself

Anthropic says the Scion footprint is about 200 times smaller in CPU and memory than the Git mirrors it replaced. It says CI and research checkouts take seconds instead of minutes, and that CI gained about 25% CPU capacity.

The migration took five months. The seven engineers also built the Antique code-review tool and about six related systems while merge volume grew tenfold. All of these figures are vendor-supplied.

One detail only the post gives: the ingestion pipeline began as a hand-written prototype, which an early internal snapshot of Claude Mythos Preview rewrote with no extra instructions. Anthropic says the rewrite used 20 times less memory and ran four times faster while keeping ordering and data-structure invariants.

Where the post tells readers to stop

The authors are blunt that Scion is not a general Git replacement. It makes sense only for enormous monorepos with tens of thousands of engineers or large agent swarms. Adopting it also means replacing code review, code search and CI, so the internal versions are bespoke and hard for others to adopt.

There is no public Scion planned. The suggested path is to apply the ideas to Git itself, such as server-side RPCs and new server architectures, while keeping the Git model.

The post also records where the agents fell short. Engineers had to reject generated patterns that broke core constraints, such as scanning commit logs linearly instead of querying the review tool. Claude's adversarial review was useful but sometimes over-engineered, and the team now reviews a change's intent rather than every line.

A global merge-queue lock did not survive agent speed either. Anthropic replaced it with post-merge and pre-deploy testing.

The offline question

On Hacker News, a Scion engineer identified as Nick G. answered a question about offline use. Willison quotes the reply: "Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do."

That is the real limit for anyone reading this as a roadmap. A system that drops offline work is a fair trade when the main user is a fleet of agents and a poor one for a laptop on a train.

Willison suspects the arrival of the Fable and Mythos model classes drove the jump in velocity, and the chart as Willison describes it marks nine public model releases from Opus 4.6 on February 5 to Fable 5.1 on September 1. That is a correlation Anthropic does not claim to have tested.

Agent-written software is also reaching outside Anthropic: our earlier coverage of DEsk, a Wayland desktop written with Claude and of REA, an agent tool with 270,097 GitHub stars covers two cases.

The post does not say whether Antique will be released or when Scion's ideas might reach a Git server. Both remain open.