Anthropic Builds Scion as AI Coding Outgrows Git

Summary

Anthropic says the rapid growth of Claude-assisted coding exposed bottlenecks in Git, whose decentralized design was created for human developers working with intermittent connectivity. At Anthropic, daily code merges reportedly rose from about 2,500 in May to 30,000 in September, or one tested and reviewed change roughly every 2.8 seconds. The company says its monorepo was nearly 100 GB, while about 25 million daily checkouts could have generated more than 600 petabytes of internal traffic without special optimization. To address the problem, a seven-person team spent five months building Scion, along with the Antique code-review, search, and permissions systems, and migrated the company’s large repository. Scion uses virtualized checkouts, streams data only when an agent needs a file or line, and removes local packaging and global locking overhead. Anthropic reports that the new system reduced CPU and memory requirements by 200 times compared with its peak Git mirror cluster, achieved cache hit rates above 99%, cut code and test fetches from minutes to seconds, and reduced continuous-integration compute by 25%. Claude also helped build the system, including rewriting a migration pipeline to use one-twentieth the memory and run four times faster. Anthropic says the project changed engineers’ work: humans focused less on typing or line-by-line review and more on architecture, intent, and stopping agents from overengineering low-probability risks. The company does not plan to commercialize Scion, but published the architecture as a warning that tools designed around human coding speed may become bottlenecks as many agents generate, test, and repair code in parallel. It expects other companies to encounter similar limits within six to twelve months, though that forecast is Anthropic’s expectation rather than an independently established result.