# Anthropic Scion: Why They Replaced Git at 30,000 Merges/Day

> Source: <https://byteiota.com/anthropic-scion-git-replacement/>
> Published: 2026-11-03 21:06:34+00:00

Anthropic hit 30,000 merged code changes in a single day and Git buckled under the pressure. So they spent five months and seven engineers building a custom source control system from scratch. The result, Scion, is not coming to a package manager near you — but the problem it solves is heading your way regardless.

## Git Was Fine — Until It Wasn’t

In May 2026, Anthropic averaged around 2,500 code changes merged per day. By September, that number was over 30,000. Claude was writing code faster than their infrastructure could process it. The monorepo’s main branch ballooned to nearly 100 GB of history. CI systems were performing approximately 25 million code checkouts per day, before counting developer environments, agent sandboxes, and research jobs. Naive git fetching at that rate would have generated over 600 petabytes of network traffic daily.

The team tried every trick in the standard Git scaling playbook: shallow clones, sparse checkouts, promisor remotes, repository snapshots, and several hundred Git mirrors running in parallel. Eventually, the workarounds became harder to maintain than the original problem. So they built something new.

## What Scion Actually Does

Scion inverts Git’s core assumption. Where Git distributes full repository copies to clients — enabling offline work — Scion puts everything on the server. Commits, rebases, and merges all happen server-side. Clients maintain bounded caches and files materialize on demand. Source control queries are exposed as APIs, so CI jobs often skip having a local repository entirely. [InfoQ’s breakdown of the architecture](https://www.infoq.com/news/2026/10/anthropic-source-control/) covers the internal mechanics in useful detail.

The unit of work is a “changeset”: an entry in a write-ahead log that stores file deltas and content-addressed blob pointers. Regional caches feed from authoritative servers backed by Spanner and cloud storage, achieving over 99% hit rates at every tier. Infrastructure footprint came in at 200 times smaller than their previous Git mirror fleet. CI checkouts dropped from minutes to seconds.

## The Part You Should Actually Care About

Scion is closed source. Anthropic is explicit that it “is by no means a suitable replacement for git for general use” and only makes sense for enormous monorepos with tens of thousands of engineers or swarms of agents. You are not getting this.

But that is the wrong lesson. The more important signal is the trajectory: 2,500 to 30,000 merges per day in four months. Anthropic notes that “the latent demand for merging code is practically infinite” and expects similar challenges industry-wide within six to twelve months. Their engineers reported a telling shift in workflow: “Very often, we found that the role of the engineers was to tell Claude ‘no’.” Coding is no longer the bottleneck. Developer infrastructure designed for human-scale output is.

## What You Can Do Today

You do not need to build Scion. [Git’s sparse checkout](https://git-scm.com/docs/git-sparse-checkout) handles early-stage scaling pressure if you use it deliberately:

```
# Sparse checkout for CI — fetch only what your job needs
git clone --filter=blob:none --sparse <repo-url>
git sparse-checkout set services/api packages/shared
```

Pair sparse checkouts with per-agent worktrees rather than a shared working copy. Add a merge queue before agent-generated PRs start conflicting at volume. Query server APIs for diff and metadata operations instead of cloning the full repository. Start tracking your daily merge rate now — the time to address this is before Git grinds your CI to a halt, not after.

[GitLab is building a Git-compatible next-generation SCM](https://about.gitlab.com/blog/scaling-git-for-the-agentic-software-factory/) targeting this exact problem: agent-native interfaces, stateless compute against shared object storage, with early tests showing workloads completing up to 50 times faster and generating up to 1,000 times less network traffic. That is probably the more realistic path for most teams — Git compatibility without the five-month infrastructure rewrite.

## The Irony in All of This

The industry spent roughly two decades moving away from centralized version control — SVN, Perforce — toward distributed systems like Git. Scion is centralized. It is always-online, server-authoritative, and explicitly trades Git’s offline-first design for server-side speed. Agents do not need offline access. Humans do. That trade-off is rational when you are running 25 million CI checkouts per day. Less obvious when you are not.

The pattern worth stealing — content-addressed storage, tiered caching, server-side operations — does not require abandoning Git. It requires treating developer infrastructure with the same rigor you would apply to production systems. Most teams are not doing that yet. The ones scaling with AI agents will need to start.