Anthropic Rebuilt Git From Scratch to Survive Agent-Scale Coding. Most Teams Don't Need a New VCS — They Need Someone to Fix the Same Bottleneck Before It Breaks Their CI.

Anthropic didn't publish this post to brag about a new tool. They published it because Git almost broke under the weight of their own coding agents.

The source

In "Redesigning source control at Anthropic," the company says daily merges in their monorepo climbed from 2,500 per day in May 2026 to over 30,000 by September 2026. The main branch history alone reached nearly 100 GB, with a compressed working copy over 25 GB, and 25 million daily CI checkouts would have generated over 600 petabytes of traffic if handled the naive way. Git's decentralized, offline-first design — built for human-scale development — became the bottleneck.

Their fix was Scion: an always-online, API-first, client-server version control system that replaced several hundred Git mirrors with a three-tier cache architecture, built and migrated in five months by a seven-person team. The results they report: a 200x smaller infrastructure footprint, checkouts two orders of magnitude faster, 25% CPU savings in CI from faster checkouts, and cache hit rates above 99% at each tier.

Anthropic is explicit that Scion isn't meant as a general Git replacement — it's built for "tens of thousands of engineers or swarms of agents prolifically writing code."

The gap that is a business

Almost no team running AI coding agents is at Anthropic's scale. But the direction of the curve is the same for anyone running more than one or two agents against a shared repo: more parallel commits, more CI checkouts, slower clones, more Git mirrors nobody is tracking. Nobody is going to build their own Scion for a 10-person team — but somebody still has to notice the repo is creeping toward the same wall and fix it with the tools that already exist.

That's a narrow, concrete audit most "AI coding agent" consultants haven't thought to offer, because they're selling agent setup, not repo infrastructure.

The money play

1. Audit a client's repo and CI the same way Anthropic's post frames the problem: merge volume trend over time, total repo size, number and cost of daily CI checkouts, and how many Git mirrors exist across the org

2. Flag the specific symptoms Anthropic named before they acted — a monorepo history creeping toward tens of gigabytes, checkouts slowing from seconds to minutes, and CI minutes scaling faster than headcount

3. Fix it with what already exists rather than building a new VCS: sparse/shallow checkouts, a scalable client like Sapling where it fits, repo-size auditing with tools like git-sizer, and consolidating redundant Git mirrors into one cached path

4. Package this as a fixed-fee "agent-scale git & CI audit" that ends with a report most teams can act on themselves, scoped to the one repo where agents commit most

5. Attach a quarterly retainer to re-check the same metrics as the client adds more coding agents, so the fix doesn't quietly get undone by the next headcount of agents

Bottom line

Anthropic had to build a new version control system because they ran out of runway on Git at agent scale. Most teams will never need to go that far — but someone still has to catch the slowdown before it's a 100 GB problem, and that's a sellable audit today.

Sources:

https://www.anthropic.com/engineering/redesigning-source-control