When Coding Stopped Being the Bottleneck
In October 2026, Anthropic's engineering blog published a post that spread fast among developers: "Redesigning source control at Anthropic." Its core claim: once AI agents write most of the code, the bottleneck moves from writing code to everything around it: source control, builds and CI.
Anthropic went from about 2,500 merged changes a day in May 2026 to over 30,000 a day by September. Git, the tool almost every team uses, couldn't keep up with their monorepo. So a seven-person team built Scion, a new version control system, and migrated in five months.
🎯 Quick Answer (30-Second Read)
- Merges grew more than tenfold in four months as AI agents did more of the coding.
- Git assumes clients work offline with a full copy of the repo. Anthropic's monorepo history is nearly 100 GB, with about 25 million checkouts a day for CI.
- Scion assumes clients are always online: the server does the heavy work, and clients fetch only what they need.
- Results: about 200x less CPU and memory than the old git mirror setup, and CI checkouts from minutes to seconds.
- Scion isn't open source, and Anthropic says it's not a general Git replacement. It wants to bring the ideas to Git platforms instead.
The Numbers That Broke Git
Anthropic ran several hundred internal git mirrors at peak to serve all that checkout traffic, plus snapshots that kept drifting behind the latest code.
Why Git Struggles Here
Git was designed in 2005 for a decentralised world: every developer has a full copy of the repository and can work offline.
Clients are humans who may be offline, so each clone is a complete, self-contained repository.
Nearly all clients are agents and CI jobs that are always online, run briefly, and need only a few files.
Workarounds like shallow clones, sparse checkouts and mirrors helped, but Anthropic described them as stopgaps. The offline-first design was working against scale.
How Scion Works
shared local object cache"] J2["Agent / CI job"] --> SD end SD --> SC["scion-cache
regional, in-memory
(99%+ hit rate)"] SC --> SS["scion-server
authoritative"] SS --> ST[("Spanner +
cloud storage")] API["Jobs using the
source-control API
(no checkout at all)"] --> SS
Anthropic's design principles:
| Principle | What it means |
|---|---|
| Clients are always online | Fetch data on demand instead of copying everything |
| The server does the heavy work | History, blame, merges and rebases happen server-side |
| Source control as an API | Many jobs query history or commit changes without any local checkout |
| Horizontal scaling | No single full copy of the repo exists anywhere; no packfile maintenance |
| Cheap, lock-free commits | Commit cost depends on the size of the change, not the repo |
| Optional, sparse working copies | Checkouts are partial or virtual when needed at all |
The unit of work is a changeset: a list of changed files relative to its parent, with content-addressed pointers to file contents. Everything else (blame, path history) is derived data, computed by one generic mechanism.
The Results
- ~200x smaller CPU and memory footprint than the previous git mirroring infrastructure
- CI and research checkouts: minutes → seconds
- 25% CPU capacity freed up in CI
- 99%+ cache hit rate at each cache tier
- Seven people built Scion, a code review tool and related systems, and migrated in five months, with Claude doing much of the implementation
Should Your Team Drop Git?
No. Anthropic says so directly: Scion "is by no means a suitable replacement for git for general use." It only makes sense for enormous monorepos or heavy agent swarms, and adopting it would mean replacing code review, search and CI too.
But the lessons apply to any team using coding agents:
- Expect PR volume to explode. Review capacity, not coding, becomes the limit.
- Agents don't need full clones. Shallow and sparse checkouts, plus API access to history, save real CI time.
- Global merge locks won't scale with many agents. Anthropic moved towards post-merge and pre-deploy testing.
- Treat developer infrastructure as a product that must scale with agent usage.
Anthropic says it hopes to work with GitHub, GitLab and the open-source community to bring ideas like server-side operations and better fetch protocols to Git itself.
Frequently Asked Questions
What is Anthropic Scion?
Scion is an in-house version control system Anthropic built to replace Git for its main monorepo. It assumes clients are always online, does history and merges on the server, and lets jobs fetch only the files they need or use an API instead of a checkout.
Is Scion open source?
No. Anthropic said Scion is bespoke to its environment and not a suitable general replacement for Git. Instead, it hopes to bring Scion's ideas to Git platforms by working with GitHub, GitLab and the open-source community.
Why did Git become a bottleneck at Anthropic?
AI agents pushed merges from about 2,500 to over 30,000 a day, while the monorepo's history grew to nearly 100 GB and CI performed around 25 million checkouts daily. Git's design, where each client keeps a full offline copy, made serving that load increasingly expensive.
Conclusion
Anthropic's Scion story shows what happens when AI agents multiply the amount of code a company ships: tools built for human-speed development become the bottleneck. Scion flips Git's offline-first design to an always-online, server-heavy model. Most teams won't need it, but every team using agents should plan for more changes, more reviews and more CI than ever before.
Sources
Last verified: 2 November 2026.
Related reads: How to Build an MCP Server in Python · Claude Code CLI Tutorial · How AI Agents Write Code Automatically · CI/CD Explained
I write about AI, developer tools and how real systems work. Follow on LinkedIn or X, or invite me to speak.