AI & DevelopmentDeveloper Tools

Anthropic Replaced Git with Scion: 30,000 Merges a Day

Abstract visualization of Anthropic Scion VCS: a central server with blue API connection lines radiating to AI agent nodes, representing agent-scale source control replacing Git
Anthropic built Scion to handle 30,000+ daily code merges driven by AI agents

Anthropic merged more than 30,000 code changes in a single day last month. In May, that number was 2,500. Git couldn’t keep up — not because something broke, but because Git was designed for humans, and Anthropic no longer writes most of its code with them. Seven engineers spent five months building a replacement called Scion. They didn’t announce it as a revolution. They announced it as a necessity.

Why Git Broke

The problem wasn’t a bug. It was a design assumption baked in from 2005: Git assumes clients work offline, need a full local copy of history, and push changes occasionally. That works beautifully for developers committing a few times a day. It does not work for swarms of AI agents committing thousands of times an hour.

By September 2026, Anthropic’s monorepo had accumulated nearly 100 GB of uncompressed history. A compressed working copy exceeded 25 GB. CI systems ran roughly 25 million checkouts daily. To manage the load, the company was running several hundred internal Git mirrors — an operationally unsustainable patchwork. Naive Git handling of 25 million daily checkouts would have generated over 600 petabytes of daily traffic. So they rebuilt from scratch.

What Scion Does Differently

Scion flips Git’s core assumptions. Where Git says “every client can serve the full repo offline,” Scion says “clients are always online, so let the server do the work.” The result is a system where commits, merges, rebases, blame queries, and history all happen server-side. Clients maintain only a bounded cache and fetch objects on demand.

The core data unit is the “changeset” — a write-ahead log entry storing file-level deltas relative to a parent, with content-addressed pointers to compressed blob storage. Caching runs in three tiers: a per-node daemon, a regional cache, and authoritative servers backed by Spanner and cloud storage. Cache hit rates exceed 99% at each tier. CI jobs can skip local checkouts entirely, querying history over an API. Read Anthropic’s full engineering post for the complete architecture breakdown.

The Numbers

  • 200x less CPU and memory than the previous Git mirroring setup
  • Checkout times dropped from minutes to seconds (two orders of magnitude)
  • 25% CI CPU capacity recovered from faster checkouts alone
  • Built by 7 engineers in 5 months, including the Antique code review tool and code search

Claude Mythos Preview contributed to the build itself — one AI-assisted rewrite produced an ingestion pipeline that consumed 20x less memory and ran 4x faster than the handwritten prototype.

What This Means for Your Team (Almost Certainly: Nothing Yet)

Anthropic is explicit: Scion “is by no means a suitable replacement for git for general use.” They will not open source it. It makes sense only for “truly enormous monorepos, with either tens of thousands of engineers or swarms of agents.” InfoQ’s technical breakdown puts the threshold at organizations running millions of daily CI checkouts before this architecture pays off.

For most teams, the practical move is simpler: use sparse checkouts with per-agent worktrees to reduce unnecessary history transfer.

git clone --filter=blob:none --sparse <repo>

Measure your checkout times before assuming you need anything more. Cursor is taking a different approach with Origin — keeping Git’s wire protocol while consolidating the full development loop. Anthropic’s own recommendation to the industry is to improve Git platforms: better server-side RPCs, improved sparse and shallow clone composition, standardized APIs across GitHub and GitLab. Developers Digest has a practical guide to what teams at different scales should actually do.

The Shift That Actually Matters

The Scion announcement is technically interesting, but the sentence buried in Anthropic’s post is the one worth sitting with: “Coding is no longer the bottleneck to building systems or doing research.”

At Anthropic, agents now generate more than 80% of the code that gets merged. Engineers describe their new role as telling Claude “no” — catching constraint violations, maintaining architectural intent, reviewing at the level of intent rather than line-by-line implementation. That’s not a distant future. That’s a seven-person team describing their current workweek.

Developer infrastructure — version control, CI, code review — was sized for a world where humans were the primary authors. That world ended at Anthropic months ago. Git probably survives this transition. The Git client model — full local clone, offline-first, human-paced — probably does not, at least not at the center of the stack. Scion is Anthropic’s answer to a problem most teams don’t have yet. But most teams are moving toward it faster than they think.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *