Anthropic Replaced Git After Claude Caused a 12x Surge in Daily Code Merges

30-Second Summary

Anthropic's AI-assisted development pipeline created a problem most engineering teams have never faced: too many code merges for Git to handle. Daily merges grew from roughly 2,500 in May 2026 to more than 30,000 by September, a 12x increase in four months. Rather than fight the constraint, a seven-person team built Scion, a purpose-built version control system, and migrated Anthropic's entire monorepo off Git in five months. Scion is not open source and Anthropic says it is not a general Git replacement, but the design ideas it shares are directly relevant to any team scaling AI coding agents.

At a Glance

• Topic: AI Infrastructure
• Company: Anthropic
• Date: October 13, 2026
• Announcement: Anthropic published an engineering post describing Scion, the custom version control system it built to replace Git internally after AI-generated code overwhelmed its existing source control.
• What Changed: Daily code merges at Anthropic grew from approximately 2,500 in May 2026 to more than 30,000 by September 2026. Scion uses server-side commits, changesets instead of traditional commits, and on-demand fetching to handle this volume.
• Why It Matters: It is the first detailed public account of a major AI lab rebuilding core developer infrastructure specifically because AI agents were writing too much code for existing tools to process.
• Who Should Care: Engineering leaders, CTOs, and technology decision-makers at companies deploying or planning to deploy AI coding agents.

Key Facts

• Daily merges at Anthropic grew from approximately 2,500 in May 2026 to more than 30,000 by September 2026 (Source: Anthropic engineering blog, October 13, 2026).
• A seven-person team built Scion and completed the migration off Git in approximately five months (Source: Anthropic engineering blog, October 13, 2026).
• Scion delivers a 200x smaller CPU and memory footprint than Anthropic's previous Git mirroring setup (Source: Anthropic engineering blog, October 13, 2026).
• Anthropic confirmed it will not open-source Scion and described it as not a suitable replacement for Git for general use (Source: Anthropic engineering blog, October 13, 2026).
• Scion uses server-side commits rather than client-side commits, and on-demand fetching rather than full repository clones.

What Happened

Anthropic published an engineering post on October 13, 2026 explaining why and how it replaced Git as its internal version control system. The driver was volume. Claude had taken over a significant share of Anthropic's internal code writing, and the resulting pace of merges exposed the limits of Git at that scale. By September, the team was processing more than 30,000 merges a day, up from around 2,500 in May.

The fundamental problem was architectural. Git was designed around the assumption that humans write code, which means commits arrive at a pace and frequency that client-server architecture can absorb. When AI agents take over code generation, that assumption breaks. Clients generating and pushing code at agent speed create a communication and synchronisation overhead that grows faster than the underlying hardware can absorb.

Scion addresses this with two key changes. First, commits are processed server-side rather than client-side, removing the round-trip cost of traditional Git push operations. Second, Scion uses changesets rather than the commit-based model Git uses, which simplifies branching and merging when many agents are working on overlapping parts of the codebase simultaneously. The result is a 200x reduction in CPU and memory footprint compared to Anthropic's previous mirroring infrastructure.

Anthropic was explicit that Scion is not intended for general use. The team stated it will not be open-sourced and that the design is specific to Anthropic's monorepo environment. What Anthropic is sharing is the design thinking: server-side commits, on-demand fetching, and changeset-based modelling are ideas that could improve Git's performance at scale if the broader community adopted compatible changes to server-side protocols and negotiation.

Why It Matters

The Git problem is coming for every serious AI team. If you are deploying AI coding agents at any meaningful scale, Git was not designed for what you are asking it to do. The constraint Anthropic hit at 30,000 merges a day is a version of the constraint your team will hit at 3,000 or 300. The threshold differs; the pattern does not.

This is the most concrete evidence yet that AI agents have crossed from assistant to primary contributor. Scion was not built because Claude occasionally writes code. It was built because Claude writes most of Anthropic's code. That shift is no longer theoretical at the leading AI labs. The question for engineering leaders is when it arrives in their own organisations, not whether.

Developer toolchain investment is now an AI strategy decision. Until now, most discussions of AI adoption focused on the AI layer: which model, which agent framework, which deployment platform. Scion shows that the toolchain one level below the AI layer matters just as much. Companies that scale AI coding agents without adjusting their underlying infrastructure will hit the same wall Anthropic hit.

The 200x efficiency gain is a signal about the scale of the problem. A 200x reduction in CPU and memory footprint is not an incremental improvement. It indicates that the old infrastructure was genuinely unsuited to the new load, not merely inefficient. Any organisation that finds its existing tooling degrading at this rate needs to treat that as an architectural problem, not a capacity problem.

The seven-person, five-month timeline is a benchmark for AI infrastructure teams. A small team rebuilt core infrastructure for one of the world's most demanding engineering environments in under half a year. That is achievable because the team had both a clear problem statement and AI assistance to accelerate the build. Organisations that have not yet invested in dedicated AI infrastructure teams should register this timeline as evidence that the investment pays back quickly.

The David and Goliath View

The Scion announcement matters more as a precedent than as a product. Anthropic is not sharing the code, and the specific engineering choices it made are tightly coupled to its own monorepo setup. What it is sharing is confirmation that the infrastructure gap between human-paced development and AI-paced development is real, large, and solvable with focused effort.

For operators running 10-to-200-person teams, the immediate takeaway is not to rebuild version control. It is to take seriously the idea that AI coding agents will change the shape of your development pipeline faster than you expect, and that the places where that change shows up first are not always the obvious ones. At Anthropic, the bottleneck was not compute, was not the model, and was not the product roadmap. It was Git.

The deeper implication is about how AI system value accumulates inside an organisation. Anthropic's ability to move fast was not limited by the quality of Claude. It was limited by the infrastructure surrounding Claude. Teams that invest in clearing those infrastructure bottlenecks compound their advantage over time, because each cleared constraint allows the AI layer to deliver more value. That logic applies to the 10-person team using AI coding tools as much as it applies to Anthropic's engineering organisation.

Where This Fits in the AI Stack

Scion operates at the developer infrastructure layer, one level below the AI model and one level above the hardware. It sits alongside CI/CD pipelines, build systems, and code review tools as part of the infrastructure that turns code into running software. The announcement is relevant to anyone thinking about the Employee Amplification Systems layer of their AI strategy: the systems that determine how much leverage each AI-assisted developer actually gets.

The D&G AI Growth Engine framework treats infrastructure investment as a prerequisite to sustainable AI scaling. Scion is a case study in what happens when an organisation runs the AI layer without proportionally investing in the infrastructure layer. The result is a ceiling, hit fast and unexpectedly.

Questions Operators Are Asking

Do we need to replace Git? Almost certainly not at your current scale. Anthropic's 30,000 daily merges is an outlier. The more relevant question is whether your current toolchain can scale to 10x your current AI-generated code volume without degrading. If you do not know the answer, that is the gap to address.

What does Scion mean for teams using GitHub or GitLab? It signals that GitHub and GitLab will need to evolve their server-side architecture to handle AI-generated commit volumes. Anthropic's engineering post explicitly calls for the broader Git community to adopt server-side improvements. Watch for GitHub and GitLab to ship capability in this direction over the next 12 to 18 months.

How do we know when we are hitting our own version of the Anthropic wall? Watch for these signals: CI queue times growing faster than headcount, merge conflict rates increasing without a corresponding increase in concurrent human contributors, and build system memory usage spiking at consistent points in the day. These are early indicators of infrastructure strain from AI-generated code volume.

Is this just relevant to software companies? No. Any company that uses software development to deliver its AI product or service faces some version of this constraint. Professional services firms using AI to accelerate document production, financial services firms using AI to build internal tools, and agencies building AI-powered client deliverables will all encounter infrastructure bottlenecks as their AI output volume scales.

What should we do now? Start tracking AI-generated code as a distinct metric. Separate it from human-written code in your CI/CD logs if you can. Set a threshold for when your current infrastructure warrants review. Seven engineers solved Anthropic's version of this problem in five months because they had a clear problem statement. The investment starts with measurement.

Citable Summary

Anthropic replaced Git with a custom version control system called Scion after Claude-generated code pushed daily merges from approximately 2,500 in May 2026 to more than 30,000 by September 2026, a 12x increase in four months (Source: Anthropic engineering blog, October 13, 2026). A seven-person team built and deployed Scion in approximately five months, achieving a 200x reduction in CPU and memory footprint compared to the previous Git mirroring infrastructure. Anthropic confirmed it will not open-source Scion and described it as not a general-purpose Git replacement, but shared its core design principles: server-side commits, changeset-based versioning, and on-demand fetching, as a model for how the broader Git ecosystem could adapt to AI-scale code generation volumes.