Anthropic Ditches Git for In-House 'Scion' as AI Agents Push Merges Past 30,000 a Day

Git, the version-control system nearly every software team on Earth relies on, has met its match: an army of AI coding agents that never sleep.

Anthropic disclosed this week that it has fully migrated its monorepo off Git and onto an in-house system called Scion, after the volume of code merged each day exploded from around 2,500 in May 2026 to more than 30,000 by September — a surge driven largely by its own Claude coding agents writing and submitting code at machine speed.

The disclosure came in an Anthropic engineering blog post published October 13. A team of just seven engineers built Scion — along with a companion code-review tool called Antique, code search and a new permissions system — and completed the full migration in about five months. Anthropic says the system will not be open-sourced.

Git is designed so that every copy of a repository is complete and works offline. That model suits most projects, but Anthropic's monorepo had outgrown it: the history of its main branch alone runs to almost 100 GB, a compressed working copy sits above 25 GB, and automated testing pulls roughly 25 million checkouts a day.

Fetched the naive way, the company says, that would mean more than 600 petabytes of Git traffic a day. At its peak, Anthropic was running several hundred internal Git mirrors just to keep up.

Standard workarounds — shallow clones, snapshots, mirrors — bought time but grew harder to run as merge volume climbed. And the offline-first design that made Git the standard was actively working against Anthropic: almost all of its heaviest users were agents and automated jobs that are always online anyway.

Scion flips Git's assumptions. Instead of every client holding a full copy and acting as its own server, it uses an always-online client-server model: commits, merges and rebases happen on the server, the repository is never fully materialized, and regional caches serve requests at over 99% hit rates.

Anthropic says the new setup runs on a roughly 200x smaller CPU and memory footprint than its old Git-mirror infrastructure, code checkouts are two orders of magnitude faster — seconds instead of minutes — and continuous-integration capacity grew by about 25%.

There is a poetic loop in the story: the company says the bottleneck is no longer writing code but “all of the developer infrastructure built for human scale: source control, builds, and CI.”

And Claude helped fix what Claude broke — engineers asked it to rewrite the Git-to-Scion ingestion pipeline, and with no further instructions it produced a version that consumed 20x less memory and ran 4x faster.

Anthropic is careful to say Scion is “by no means a suitable replacement for git for general use”: it needs its own code review, code search and CI tooling, and was built for an extreme, agent-heavy workload few companies have.

But the company plans to share its approach with GitHub, GitLab and open-source projects, and it suspects others using AI coding tools will feel the same pressure soon enough. The message for the industry: when your biggest committer is a machine, your tooling has to grow up fast.