Anthropic replaced Git inside its monorepo with Scion, an in-house version-control system, after daily merged code changes climbed from around 2,500 in May 2026 to more than 30,000 by September 2026.
A seven-person team built Scion, a companion code-review tool called Antique.
And several related systems, then finished the full migration in five months (Anthropic, Digg).
Scion moves commits, merges, rebases, file history. And blame onto a server, keeps clients on a bounded local cache. And exposes source control through an API so many CI jobs no longer need a local repository at all. The reported payoff: a 200x smaller CPU and memory footprint than the old Git mirroring infrastructure. And CI checkouts that dropped from minutes to seconds (bmannconsulting notes). One caveat comes from Anthropic itself: Scion is not a general replacement for Git.
That’s the whole announcement in one paragraph. The interesting part for the rest of us is what it reveals about where the bottleneck in software moved.
Why Git Broke at Anthropic
Git didn’t break because it’s bad software. It broke since Anthropic’s workload changed shape underneath it. In May 2026 the monorepo averaged around 2,500 merged changes per day. And by September that number was over 30,000 (Anthropic).
Anthropic says its developer infrastructure, built for human rather than agent scale, became the bottleneck as that growth ran past it (Digg).
The surrounding scale is hard to picture. Anthropic says its continuous-integration systems perform roughly 25 million code checkouts each day, before counting developer environments, agent sandboxes, and research jobs (Digg). Every one of those checkouts was pulling against infrastructure that assumed a modest number of human-driven clients.
Here’s the part I keep chewing on as someone who runs coding agents against my own repos daily. The agents didn’t get slower and the humans didn’t get faster. Nothing about any single commit changed. What changed was volume. And volume is exactly the axis where a system designed around complete local repositories hurts most.
When nearly all of your clients are agents and jobs rather than people, Anthropic concluded, the design should assume a server that is always reachable and let clients fetch data only when they need it (Anthropic).
How Scion Works: The Server Does the Thinking
Scion inverts Git’s core trade. Git gives every client a full copy of the repository and does the work locally. Scion puts the work on the server and gives clients a thin slice. The specifics from Anthropic’s engineering post and coverage:
– The server answers file-history and blame queries, creates commits, and performs merges and rebases (Anthropic).
– Clients keep a bounded local cache and fetch only the trees and files they need, while sparse and virtualized checkouts avoid materializing the full repository (Digg).
– Instead of Git commits pointing to complete trees, Scion records a changeset as the files changed relative to its first parent, with content-addressed pointers to compressed file data (Digg).
– All the source-control queries a client uses are also exposed via API, so many jobs don’t need a local repository at all (Anthropic).
– Repository data is distributed, every component is independently scalable, and there is no fully materialized copy of the repo anywhere, even on the server (Anthropic).
Notice what got traded away.
Scion assumes clients can reach the server whenever they need data, which is fine when your clients are agents on managed infrastructure and anathema to Git’s offline-first promise.
If the server is unreachable, the agent’s next fetch waits. That’s a rational bet for a lab where nearly every client is a machine with a network.
And a bad bet for a laptop on a plane.
Anthropic didn’t lose anything it valued, given that its agents never worked offline in the first place.
The Numbers That Justified the Rewrite
The migration story is small enough to be almost insulting to enterprise planning cycles.
One seven-person team built Scion, the Antique code-review tool. And several related systems, then completed the entire monorepo migration in five months (Digg). Five months, seven people, the version-control backbone of a frontier lab.
The performance claims explain the urgency.
Anthropic says Scion’s infrastructure footprint is 200x smaller than its previous Git mirroring infrastructure in CPU and memory terms (bmannconsulting notes). CI and research jobs check out code two orders of magnitude faster, in seconds rather than minutes.
And CI alone realized a 25% CPU capacity win (bmannconsulting notes).
That 25% figure is the quiet headline. A capacity win means machines you already paid for suddenly do more work, which at 25 million daily checkouts is an enormous amount of recovered compute. When people quote this story as “Git is dead,” they’re missing that the actual result was an infrastructure bill cut by two orders of magnitude.
You Are Not Anthropic, and That Is the Point
Anthropic says it plainly: Scion is not a general replacement for Git. And its design is aimed at exceptionally large monorepos and workloads produced by large numbers of coding agents (Digg). If you run a small business with a handful of repos and a couple of agents, your daily merge count is a rounding error next to 30,000. Do not rewrite your version control. Anyone telling you Git is dying for teams your size is selling something.
But two things here deserve your attention anyway. First, the trend line, not the tool. Anthropic’s merges grew more than tenfold in four months. And the wall it hit was checkouts, CI, and review capacity, not code generation. If agents write most of your code next year, your constraints move to the same places, just at lower volumes. The leading indicators are measurable today: daily merge count, average CI checkout time. And how long changes wait in review. Track them before they hurt.
Second, apply some skepticism to the merge number itself.
Thirty thousand merged changes per day is partly an infrastructure problem and partly a behavior problem, as agents merge small changes constantly in a way disciplined humans don’t. Before you conclude you need new infrastructure, ask whether your agents need to merge that often at all. Batching agent work and gating merges costs nothing and often dissolves the bottleneck. Anthropic’s scale made the rewrite unavoidable, and nothing about your scale does.
What to Do This Week
Pull three numbers from your own setup: yesterday’s merged-change count, your average CI checkout time. And your worst review-wait over the past month. Write them down somewhere you’ll see them again. That baseline is exactly what I capture first when I wire agents into a build pipeline, since the trend in those numbers tells you months in advance whether you’re drifting toward Anthropic’s problem or just running hot. Then read Anthropic’s engineering post yourself, since it’s the rare announcement where the primary source beats the commentary (Anthropic). Git will carry you for a long time yet. And now you know the number at which it stopped carrying them.
