Anthropic ditches Git for custom agent-era version control

Everyone is debating whether AI will replace engineers, but at Anthropic, Claude just broke their developer tools instead.

In May 2026, Anthropic averaged roughly 2,500 merged code changes per day. By September, that number skyrocketed past 30,000 per day. As Claude autonomously iterated on research loops and built services from scratch, code output spiked by 12x, and the bottleneck shifted completely from writing software to managing it.

Anthropic runs a massive monorepo where the main branch history approaches 100 GB and a working copy tops 25 GB. Between research sandboxes and CI jobs running roughly 25 million checkouts a day, standard Git operations would have generated an impossible 600 petabytes of daily network traffic.

Rebuilding source control for always-online agents

To survive, Anthropic previously ran hundreds of internal Git mirrors and relied on sparse checkouts. But as merge rates exploded, the mirrors fell behind, disk limits cracked, and Git's foundational design simply gave out.

Git was built around a basic premise: human developers work offline, so every local clone must be a full, standalone repository. But nearly all of Anthropic's daily clients are agents and CI runners that are online by definition.

So Anthropic asked a simple question: what would version control look like if you assumed clients never go offline? The answer is Scion, a custom version control system built and deployed across their entire monorepo in just five months. In classic full-circle fashion, Claude wrote a huge chunk of Scion's infrastructure.

Scion flips Git on its head. It is entirely server-side and horizontally scalable, meaning there is no full materialized copy of the repository anywhere. Commits are lock-free changesets in a write-ahead log. Clients don't even need full checkouts; files are virtualized on demand, and automated jobs can query diffs and blame directly through APIs without touching local disks. Alongside Scion, Anthropic built a new review tool called Antique, plus internal code search and access control.

Why it matters

Anthropic's migration offers an early look at what happens when agentic software engineering meets legacy developer infrastructure.

For regular developers, Git isn't going anywhere anytime soon. You don't produce 30,000 merges a day, and Scion is an internal system tailored specifically to Anthropic's extreme environment. But if you run an engineering team leaning heavily into AI coding agents, this is a clear preview of the wall you will eventually hit.

The broader impact hits platforms like GitHub, GitLab, and Atlassian. For twenty years, developer tooling was optimized for human reading speeds, local filesystems, and occasional manual pushes. When the primary author of your codebase is a tireless agent creating millions of micro-tasks, traditional CI and decentralization become massive liabilities.

We think Anthropic made a brilliant call by ditching Git's offline dogma instead of endlessly piling mirrors onto a broken foundation. That said, it creates a massive dependency on server availability and distributed databases like Spanner. If that centralized backbone stumbles, development halts instantly, with zero local fallback for the humans left in the loop.