Git wasn't built for a world where your AI writes more code than your entire engineering team.
The Summary
- Anthropic built Scion, a custom source control system to replace Git, because Claude is generating code faster than Git can handle
- Daily code merges at Anthropic have jumped 10x since May as AI agents increasingly write production code
- The shift signals a fundamental infrastructure problem: legacy developer tools weren't designed for machine-speed code generation
The Signal
Anthropic didn't replace Git on a whim. When your AI coding assistant starts writing code at volumes that make human pull requests look quaint, version control becomes the bottleneck. The company's daily code merges have increased tenfold since May, a pace that Git's 20-year-old architecture struggles to match.
Scion isn't just faster Git. It's purpose-built for a development workflow where Claude generates thousands of code changes daily, each requiring review, merging, and conflict resolution at speeds human-focused tools never anticipated. The infrastructure layer of software development is being rewritten because the producers of that software are no longer primarily human.
"Git wasn't built for a world where your AI writes more code than your entire engineering team."
This matters beyond Anthropic's walls. Every company building serious AI coding tools will hit this ceiling. GitHub, GitLab, and others still optimize for human commit patterns: dozens or hundreds of merges per day, not thousands. When agents become the primary code authors, the entire stack beneath them needs rethinking.
Three implications cascade from this:
- Developer tooling companies face an existential rebuild-or-die moment
- Code review processes need automation layers humans never required
- Software security gets harder when you're auditing machine-generated code at machine scale
The shift highlights growing infrastructure needs as AI-driven coding accelerates. But it also exposes a risk: custom tools create lock-in and reduce transparency. When your version control is proprietary, auditing what your AI actually ships gets murkier. That's fine for Anthropic's internal systems. It's a problem if this pattern spreads industry-wide without open alternatives.
The Implication
If you're building agent-based development tools, your version control strategy is now a competitive moat. The companies that solve machine-speed code management will capture the agent economy's most valuable workflow: writing software. Watch for Git alternatives to emerge as open-source projects or venture-funded startups in the next 12 months.
For developers, the message is clear: the tools you know are being rebuilt around you. The shift from human-centric to agent-centric infrastructure is happening faster than most realize. Learn how AI coding workflows actually operate at scale, because the merge queue is about to look very different.