Anthropic Ditches Git: How 30,000 Daily AI Merges Forced a New Version Control System

What Happened: A 12x Merge Surge Broke Git

Anthropic revealed this week that it has moved its entire monorepo off Git and onto Scion, a version control system it built in-house — because AI coding agents broke the old one. Daily code merges at the company climbed from roughly 2,500 in May 2026 to more than 30,000 by October, a 12x surge driven overwhelmingly by Claude writing, reviewing, and shipping code alongside human engineers.

Git was designed in 2005 for human-paced development: dozens of developers, a few merges a day, conflicts resolved over coffee. It was never designed for fleets of AI agents committing machine-paced changes around the clock. At 30,000 merges a day, operations that took seconds started taking minutes, merge queues backed up, and the overhead of Git's object model became a genuine engineering bottleneck — not an inconvenience, a ceiling on how much agentic code the company could absorb.

The number itself is the story. Thirty thousand daily merges is not "developers moving faster." It's a categorical shift: the majority of code changes at one of the world's most valuable AI labs are now machine-generated, and the tooling layer underneath had to be rebuilt to match.

Scion: Version Control Built for Machines

Scion is Anthropic's answer — a version control system designed from the ground up for repositories where agents, not humans, are the primary committers. According to the company's disclosures, the migration took the entire codebase off Git in about five months, and Scion now handles the full weight of Anthropic's engineering output with a smaller infrastructure footprint, with operations that once took minutes completing in seconds.

While full technical details are still emerging, the design goals are clear from the problem it solves:

• Machine-paced concurrency: thousands of simultaneous change streams without Git's locking and rebase contention.
• Cheaper history: storing tens of thousands of daily changes without the object-store bloat that plagues large Git monorepos.
• Agent-native semantics: first-class support for the commit-review-merge loops that coding agents run continuously.

Notably, Anthropic says Scion will not be open-sourced, at least for now. It's internal infrastructure built to keep pace with the company's own agentic coding pipeline — a competitive tool, not a community project.

Built With Claude, by Seven Engineers, in Five Months

The most telling detail: Anthropic says Claude itself helped build Scion. A team of roughly seven engineers shipped a Git replacement — a category of software that took the industry a decade to standardize on — in about five months, with AI doing a substantial share of the heavy lifting.

That's the recursive loop of 2026 in miniature: AI agents generated so much code that version control had to be reinvented, and AI agents helped reinvent it. If you needed proof that agentic coding tools have crossed from "autocomplete on steroids" to "systems that rebuild core infrastructure," this is it.

It also tracks with what other frontier labs are reporting. Google says Gemini 4 Argon helped its teams separate a six-million-line legacy folder inside Chromium's 40-million-line codebase, turning months of refactoring into weeks. The pattern is consistent everywhere: agent-written code volume is growing faster than the human-era tooling underneath it can absorb.

Antique: The Code Review Bot for AI-Sized Pull Requests

Scion isn't alone. The same team of seven also built Antique, a code review tool designed to handle the surge of AI-generated changes. Human reviewers can't meaningfully eyeball 30,000 merges a day — review has to become an automated, policy-driven layer that filters what reaches human eyes.

This is the quieter half of the story, and arguably the more important one for everyone else. The bottleneck in AI-assisted development is no longer writing code. It's absorbing code: reviewing it, trusting it, and merging it safely. Whoever solves automated review at scale solves the real problem.

What It Means for Your AI Coding Stack

You're probably not running a 30,000-merge monorepo. But the forces that pushed Anthropic off Git will reach ordinary teams in three concrete ways:

• CI/CD costs will spike before anything else. Agent-generated pull requests multiply pipeline runs. Teams that don't batch, dedupe, or queue agent work will drown in compute bills and queue times.
• Review becomes the constraint. If agents write five times more code, human review capacity becomes the choke point. Automated review tooling — what Anthropic built Antique for — moves from nice-to-have to mandatory.
• Version control tooling will diversify. Git stays fine for most teams for years. But the era of assuming Git is the permanent, final answer to version control is over. Expect vendor and open-source responses to Scion's existence.

The practical takeaway: if you're deploying AI coding agents across a team, measure your merge and review throughput now. That's the curve that breaks first.

AI Coding Tools That Scale With Agent Traffic

The tools driving this merge surge are the same ones available to every developer. If you're building an agentic coding workflow in 2026, these are the categories worth watching on aitrove.ai's AI coding tools directory:

• Agentic coding tools like Claude Code, Cursor, and GitHub Copilot — the engines generating this volume of changes.
• Multi-agent orchestration platforms that run parallel agents on separate worktrees, mirroring the concurrency that overwhelmed Git.
• AI code review and security scanning tools — the "Antique category" — which will define how safely teams can absorb agent output.

Compare options, pricing, and real pros and cons in our curated best AI code assistants guide before committing your workflow to any single vendor.

Frequently Asked Questions