Claude generates so much code that Anthropic had to rebuild its version control system

Anthropic is replacing Git with Scion, its own code management system, after surpassing 30,000 merged changes per day thanks to Claude.

At Anthropic, writing code has become faster than managing it. The heavy use of Claude has driven an explosion in internal development volume, to the point of making traditional code management tools difficult to maintain at this scale. To solve this problem, the company developed Scion, a version control system designed entirely for AI agents, and migrated its entire main repository off Git in just five months. From 2,500 to over 30,000 changes per day

The progression gives an idea of the change in scale. In May 2026, Anthropic was merging around 2,500 code changes per day. By September, that volume exceeded 30,000—a twelvefold increase in four months.

This output relies heavily on Claude, which is used to develop new services, optimize programs, and conduct research experiments. Therefore, the bottleneck no longer comes primarily from writing code, but from the infrastructure tasked with recording, checking, and integrating it into existing projects.

Anthropic's main repository represents nearly 100 GB of Git history and over 25 GB for a compressed working copy. Added to this are roughly 25 million code fetch operations every day for automated testing, not counting developer and agent environments. Scion, a system designed for always-connected agents

Git was designed to allow every developer to work with an autonomous copy of their repository, including offline. This architecture remains relevant for most projects, but it becomes expensive when thousands of agents and automated processes continuously access the same set of files.

Scion reverses this logic. The system centralizes complex operations on servers and transmits only the necessary data to clients. Agents can consult the history, fetch changes, or query the repository directly via API, without having to download a full copy.

Commits, merges, and reorderings of changes are executed server-side. Files are distributed across an infrastructure capable of scaling its capacity independently as needed, while local copies remain limited to the elements actually in use.

According to Anthropic, Scion requires 200 times fewer CPU and memory resources than its former Git mirror infrastructure. Code fetches used by tests and research tasks are now about 100 times faster, dropping from several minutes to a few seconds. Continuous integration tests also benefited from a 25% gain in CPU capacity. Seven engineers and Claude to rebuild the infrastructure

The development of Scion and Antique, the accompanying internal code review tool, engaged a team of just seven people over five months.

Claude participated extensively in their design. Anthropic notably cites a first version of its Git migration system, initially written by hand, then entirely reworked by an internal version of Claude Mythos Preview. The result consumed 20 times less memory and ran four times faster.

The group nevertheless insists on the role of the engineers, responsible for defining the architecture, arbitrating constraints, and correcting inappropriate decisions made by the agents. Human reviews remain systematic, but focus more on the intent and consequences of the changes than on the exhaustive reading of every line. A special case that could foreshadow other changes

Anthropic does not plan to market Scion as a universal alternative to Git. Its architecture addresses the constraints of a gigantic repository fueled by an exceptionally intensive development activity, hardly comparable to that of most companies.

This experience nonetheless raises a broader question: what becomes of software development tools when the quantity of code produced far exceeds the usual capacity of human teams to integrate it?

Anthropic is already beginning to revise its testing, validation, and monitoring methods. With Scion, the company is documenting a concrete consequence of the massive adoption of programming agents: infrastructure designed to support developers must now absorb the work of machines capable of continuously producing code.