Anthropic Rebuilds Source Control for Agents
Executive Summary
AI-generated code is forcing Anthropic to redesign the infrastructure that receives, reviews, and tests it. The company says daily merged changes rose from roughly 2,500 in May to more than 30,000 in September, making source control its biggest development bottleneck. Its response was Scion, an in-house version-control system built around always-online agents rather than developers carrying complete local repositories.
The consequential claim is not simply that Claude writes code faster. It is that higher output changes the economics—and the safety assumptions—of the entire development lifecycle. Anthropic’s engineering account provides unusually concrete evidence of that shift, while explicitly warning that its workload and internal token budget are exceptional. This is a case study in agent-scale engineering, not a general recommendation to abandon Git.
What Happened
In its October 13 engineering post, Anthropic describes a seven-person team building Scion, the Antique code-review tool, and related systems, then fully migrating its monorepo from Git in five months.
The reported workload explains the urgency: almost 100 GB of main-branch history, a compressed working copy exceeding 25 GB, and approximately 25 million CI checkouts daily. Naively fetching those working copies would imply more than 600 PB of daily traffic—not the company’s measured traffic, but an illustration of why conventional replication becomes untenable. Anthropic had already used partial checkouts, snapshots, and several hundred internal Git mirrors.
Scion moves commits, merges, rebases, history, and blame queries onto servers. Clients keep bounded caches and request only the files or metadata they need; some jobs operate entirely through APIs without materializing a working copy. This sacrifices Git’s self-contained offline model in favor of sparse access and horizontally scalable services.
Anthropic reports a source-control infrastructure footprint 200 times smaller in CPU and memory requirements than its previous mirroring setup. CI and research checkouts became two orders of magnitude faster, and CI recovered 25% of CPU capacity. These are substantial first-party operational claims, not independently reproduced benchmarks.
Why It Matters
The most revealing change is downstream of code generation. Anthropic says a conventional globally locked merge queue no longer works effectively at its scale: hundreds or thousands of changes can land while tests are running. It has instead invested in post-merge and pre-deployment testing, accepting that concurrent changes can conflict and relying on Claude to help repair them quickly.
That is a different risk allocation, not merely a faster implementation of the old workflow. Rapid repair does not prove that every conflict is harmless. The relevant question becomes how reliably a system detects and contains problems before deployment, rather than whether each change passed an isolated pre-merge gate.
The human role also remains concrete. Anthropic says engineers review every change request, focusing on intent rather than reading every line. They choose architectural trade-offs and reject agent solutions that violate system constraints. One example involved agents scanning commit history linearly to identify when changes merged, instead of querying the review system. Adversarial AI review sometimes introduced excessive validation and complexity, requiring engineers to push back.
More generated scrutiny, in other words, does not automatically produce better engineering judgment.
The Bigger Story
This reinforces a recurring distinction: the ability to complete a task does not establish the ability to respect its constraints. Here, the problem appears in architecture and scalability as well as permissions. Agents can produce a locally plausible solution that undermines the larger system; experienced engineers still decide which problems deserve solving and which approaches must be ruled out.
It also complicates simple productivity narratives. Merged-change counts measure activity, not customer value, defect rates, or net economic return. Anthropic’s unusually generous access to tokens limits how directly other organizations can extrapolate its results. And Simon Willison’s commentary correctly distinguishes public model-release dates from internal availability: a chart overlay alone cannot establish which release caused the acceleration.
The transferable lesson is narrower and stronger than “replace Git”: when code becomes cheap to produce, coordination, verification, and data movement can become the expensive parts. Scion makes that bottleneck shift tangible.
Further Reading
- Redesigning source control at Anthropic: detailed architecture, reported performance gains, and explicit limits on where the design makes sense.