Dating note: Anthropic’s engineering post was reported on October 14. I haven’t read the original, so this is based on AI Weekly and DevOps.com. All numbers are Anthropic’s, as relayed.

Why it exists

Anthropic says daily merges rose from about 2,500 in May 2026 to more than 30,000 by September, mostly from agent-written code. At that volume, Git’s distributed design struggled:

  • Main-branch history is close to 100 GB, and a compressed working copy is over 25 GB.
  • CI alone runs about 25 million checkouts a day. Naive fetching would mean over 600 PB of Git traffic daily.
  • The old setup peaked at several hundred Git mirrors.

What Scion is

A custom version-control system built by a team of seven engineers over about five months. As DevOps.com describes the design:

  • A server handles commits, merges, rebases, blame and file history. Clients keep a bounded local cache, and many CI jobs query the server through an API rather than keeping a full copy.
  • Commits are stored as changesets, so cost scales with the change, not the repo.
  • Branches and tags are discouraged because global state is hard to scale.
  • Checkouts are sparse or virtualized, so jobs pull only what they read.
  • Data is content-addressed and immutable, with three tiers (a per-node daemon, regional caches and authoritative servers backed by Spanner and cloud storage). Anthropic reports cache hit rates above 99% per tier and a 200x smaller CPU and memory footprint than the old mirror setup, with checkouts in seconds.
  • A companion review tool, Antique, lets reviewers focus on each change’s intent instead of reading every line.

Claude’s role

Anthropic says Claude wrote most of the code. In one example, an early internal build of a Claude Mythos model rewrote the ingestion pipeline, reportedly using 20x less memory and running 4x faster. Engineers say they still own system design and trade-offs.

The caveats Anthropic itself gives

The headline “replaces Git” is not what Anthropic claims:

  • Scion only makes sense for very large monorepos, such as tens of thousands of engineers or large agent swarms.
  • It requires its own review, search and CI systems.
  • Anthropic says it is not a suitable replacement for Git for general use and won’t open-source it. One summary’s original headline said otherwise, but its body contradicted that.
  • Anthropic expects more teams to hit similar limits in six to 12 months, though most won’t build their own system. It wants to push ideas back to GitHub, GitLab and the open-source community: richer server-side forge APIs and better sparse and shallow checkouts.
  • A global merge queue doesn’t scale at agent volume, so it leans on automated post-merge and pre-deploy testing.

What I take from it

I’d draw three lessons, mine and not Anthropic’s:

  1. Agents change the shape of load before they change quality. The first thing to break was infrastructure, not code.
  2. Review becomes the bottleneck. Antique’s “review the intent” idea is the same pressure I noted in the Claude Code changelog post, where hooks are being used as guardrails.
  3. Track your own merge volume. Analyst Mitch Ashley of The Futurum Group advised teams adding coding agents to watch merge volume and plan for testing, rollback and approval records after merges.

If you only run a normal-sized repo, the practical takeaway is smaller: worktrees per agent, as in the Codex worktrees post, plus tests that run before anything lands.

Sources: AI Weekly, DevOps.com.