Why Git Is Breaking Down: Inside Anthropic's Leap to 30,000 AI Code Merges a Day
The landscape of software development is undergoing a massive paradigm shift. For decades, the primary bottleneck in engineering has been human typing speed, debugging cycles, and the cognitive overhead of writing code line by line. However, recent milestones achieved by AI labs suggest that coding is no longer the bottleneck.
The AI Coding Explosion: From 2,500 to 30,000 Merges
In May 2026, Anthropic's internal monorepo recorded roughly 2,500 daily code merges. Fast forward to September 2026, and that number skyrocketed to over 30,000 merges in a single day. This exponential surge is primarily driven by Claude coding agents autonomously writing, testing, and merging code at unprecedented speeds.
While this productivity boom is a triumph for AI-assisted engineering, it triggered an unexpected infrastructure crisis. Software engineering teams heavily utilizing AI agents are facing severe version control and continuous integration (CI) scalability walls that legacy tooling simply cannot handle.
Why Traditional Git Fails at Agent Scale
To understand why traditional version control systems break down under agentic workflows, we have to look at the numbers. Anthropic's monorepo main branch history approaches 100 GB, working copies exceed 25 GB, and continuous integration executes roughly 25 million checkouts daily.
Standard, human-centric version control systems like Git were never architected for thousands of autonomous non-human actors submitting changes simultaneously. When deployed at this scale, Git-mirroring setups suffer from massive resource bloat, dragging CI checkouts out into frustrating multi-minute waits. Industry analysts and engineering experts point out that code generation is no longer the primary constraint; developer infrastructure built for human scale has become the new bottleneck.
Meet Scion and Antique: Anthropic's Custom Infrastructure Cure
Faced with this scaling wall, Anthropic took a radical step. A small team of seven engineers completely migrated off Git over a period of five months, building a custom version control system named Scion alongside a specialized code review tool named Antique.
Scion is specifically optimized for agent-centric workflows. It reduced CPU and memory footprint by an astounding 200x compared to previous Git-mirroring setups and accelerated CI checkouts from minutes down to mere seconds. Despite its massive success, Anthropic has stated that Scion will not be open-sourced and is not intended as a general-purpose replacement for Git. It is a bespoke solution tailored to extreme AI-native internal scaling.
The New Bottleneck: Human Governance and Infrastructure Realities
This milestone has sparked intense debate across online communities like Hacker News and Reddit. While developers expressed astonishment, they also voiced deep skepticism regarding how code quality and architectural integrity can possibly be maintained across 30,000 daily merges. Many engineers joked that a software developer's primary job in the agent era has evolved into telling Claude 'no' when it attempts to break system limits.
Tech commentators predict that other engineering organizations scaling up AI agents will face similar infrastructure walls within 6 to 12 months. As we move forward, the challenge shifts from generating code to governing it. Before adopting heavy AI coding workflows, engineering teams should carefully evaluate their internal development pipelines and CI systems for potential bottlenecks.
Frequently Asked Questions
How does Anthropic practically handle merge conflicts and prevent agentic regression or noise at a scale of 30,000 daily merges without drowning human reviewers?
While automated testing pipelines vet the bulk of agent-submitted code, human oversight relies heavily on customized review tools like Antique designed to filter out noise, handle high-frequency concurrent branches, and flag potential architectural drifts before they destabilize the main branch.
What specific architectural and networking protocols does Scion use under the hood to achieve its reported 200x resource footprint reduction?
Anthropic designed Scion from the ground up to bypass the overhead of decentralized graph histories used by Git, implementing a server-centric model optimized for rapid snapshot retrieval and minimal local storage replication during massive parallel CI checkouts.
