In May 2026 Anthropic was incorporating about 2,500 code changes each day, and four months later code changes had topped 30,000, prompting the company to build its own version control system, dubbed Scion.
The software produced by its AI coding agents was developed so rapidly that the infrastructure responsible for managing their work could hardly keep up.
Rather than continuing to expand its Git infrastructure, Anthropic developed and migrated internal monorepo to Scion in just five months. The replacement uses 200 times less CPU and memory than its previous Git-mirroring infrastructure, while freeing up 25% of continuous integration (CI) CPU capacity (Anthropic, 2026).
Perhaps the more interesting development is what Anthropic's experience reveals about the limits of developer infrastructure designed around human activity. Git was designed around distributed development, allowing engineers to maintain repository history and work offline. Anthropic's environment looks quite different. Most repository clients are now online agents, automated jobs, and CI systems, rather than developers working locally. Its main branch history approaches 100 GB, while CI alone performs approximately 25 million code checkouts daily. Downloading complete working copies repeatedly would create enormous infrastructure needs.
Anthropic initially relied on familiar optimizations, including sparse checkouts, shallow clones, and internal Git mirrors. Eventually, managing hundreds of mirrors became part of the problem. Scion reverses Git's traditional architecture by moving much of the work from individual clients to centralized services.
Commits, merges, history queries, and other operations run server-side. Clients retrieve only the files they need, while layered caches reduce repeated data transfers. Instead of requiring every CI job to download repository data to inspect a change, Scion can answer many queries directly through APIs. Checkouts that previously took minutes now take seconds, according to Anthropic.
Scion was itself largely built with AI. Working with Claude, a team of just seven engineers developed the version control system, a separate code review tool called Antique, and supporting infrastructure in five months. Claude wrote most of the code, while the engineers remained responsible for the architecture, technical trade-offs, and identifying approaches that looked reasonable but would struggle at scale.
Claude may have written most of Scion's code, but the engineers remained responsible for its architecture and technical trade-offs. Anthropic describes instances where agents proposed inefficient solutions that required human involvement. Rather than making engineering judgment less important, writing code faster has made decisions about architecture and system scalability more consequential.
The same pressure is reshaping Anthropic's development pipeline. Traditional merge queues become difficult to manage when thousands of changes arrive while earlier tests are still running. Anthropic is responding with greater reliance on post-merge testing, pre-deployment validation, and automated rollback mechanisms.
Scion isn't intended to replace Git for most developers. Anthropic's enormous repository and unusually high volume of agent-generated changes create demands few organizations currently face.
Yet the experience provides an early sign of the direction in which AI-assisted development might be going. As agents take away the limitations on code generation, the bottlenecks move to source control, testing, and deployment. There will be little benefit in producing ten times more code if the infrastructure responsible for verifying and releasing it is unable to keep up. It all comes down to how much change their systems can safely handle.