On October 13, Anthropic published a post detailing how it replaced Git with an in-house
version control system called Scion across its entire internal monorepo. The company says the
growth in code volume generated by its own Claude agents eventually overwhelmed classic Git
infrastructure, built for human developers rather than thousands of automated, always-connected
clients.
A twelvefold jump in changes over four months
The number of changes merged into the monorepo each day rose from roughly 2,500 in May 2026 to
over 30,000 by September. The main branch’s history weighs close to 100 GB, and a compressed
working copy exceeds 25 GB. Continuous integration checks alone amounted to about 25 million
daily checkouts, before even counting developer environments, agent sandboxes and research
jobs.
Why Git stopped keeping up
Git’s model relies on a full copy of the history on every clone, a reasonable choice when
clients are developer machines connecting occasionally. At Anthropic, nearly all clients are
agents or continuous-integration jobs that stay online permanently: fetching history naively
would have meant over 600 petabytes of Git traffic per day. To compensate, the company ran
several hundred internal Git mirrors, an infrastructure that itself became the main cost
center.
| Metric | Before (Git mirrors) | With Scion |
|---|---|---|
| Changes merged / day | ~2,500 (May 2026) | >30,000 (September 2026) |
| Infrastructure CPU/memory footprint | Baseline | ~200x lower |
| CI checkout duration | Several minutes | Seconds |
| CPU capacity available for CI | Baseline | +25% |
| Cache hit rate (per tier) | Variable | >99% |
Scion and Antique: the chosen architecture
Scion replaces Git’s model with a write-ahead log: each changeset describes files
changed relative to its first parent and points to content-addressed blobs, with no pack files
and no ref locking. Derived data (blame, path history, commit graph) is computed recursively from
parent changesets, and clients only fetch the trees and blobs they actually need — which makes
partial checkouts native, with no special configuration required. Three serving tiers (a local
daemon, regional caches, authoritative servers backed by Spanner) each maintain a cache hit rate
above 99%. Antique, the accompanying code review tool, rounds out the system with code search and
fine-grained, per-source permissions.
Classic Git history:
full clone -> ref locking -> packfiles -> expensive fetch for an agent
Scion model:
changeset (write-ahead log) -> content-addressed blobs -> partial fetch
via tiers (local daemon / regional cache / Spanner server), >99% hit rate
“The bottleneck, now, is all of the developer infrastructure built for human scale.”
Scion is not a public product or a general-purpose Git replacement: Anthropic says so
explicitly in its post. The relevance for ordinary web teams remains indirect, but it illustrates
a tension that will reach other tool vendors as agent-driven code pipelines generate a volume of
changes that classic merge queues were never designed to absorb.
A seven-person team, five months of migration
Building Scion, Antique and about half a dozen related systems took a team of seven over five
months, with Claude itself writing a large share of the migration code. The engineers describe
their role as mostly telling Claude no when it tried to over-engineer a solution, and reviewing
the intent of changes rather than every line. The post goes as far as stating that every engineer
who worked on the project now rates Claude as better than them at writing code — a claim worth
treating with the usual caution for posts published by the model’s own maker, but one that echoes
similar reports elsewhere in the ecosystem, including around the MCP protocol and its rapid
adoption across agentic tooling chains.
Key takeaways
Anthropic’s case remains an extreme one, specific to a company running its own agents at very
large scale on its own code. It still offers a useful signal: version control and continuous
integration tooling designed for occasionally connected humans is starting to show its limits
against permanently automated clients, and server-centric alternatives to Git will likely gain
visibility in the months ahead, without replacing Git for the vast majority of teams.
I remain skeptical of the practical relevance of this kind of announcement for most of the
teams I work with: none of them come anywhere close to the 30,000 daily changes that drove this
project. What does seem worth watching, though, is the growing pressure coding agents put on
ordinary CI pipelines, even at a much more modest scale — a concern that already applies to the
projects I’m involved in. — Simon Janvier
Further reading: the full post is available on
Anthropic’s engineering blog.
