Anthropic revealed that in May this year, its system was handling an average of approximately 2,500 code merges per day. By September, this figure had skyrocketed to over 30,000 merges daily. Such rapid growth put a strain on existing development tools, including the Git version control system, which struggled to keep pace. Initially, the team implemented three temporary fixes for the continuous integration service, but these solutions proved increasingly inadequate over time. Eventually, an engineer undertook a major refactoring of the service. To tackle the bottleneck in source code management, Anthropic formed a seven-member team dedicated to developing a new version control system, Scion, specifically tailored for ultra-large-scale code repositories. This effort spanned five months. The company successfully migrated its entire monolithic codebase from Git to Scion. Scion operates on a model where the server handles operations and the client retrieves data as needed, incorporating hierarchical caching mechanisms. As a result, Anthropic’s source code management infrastructure is now just 1/200th the size of its previous Git mirror setup. During the development process, the team leveraged Claude to rewrite the prototype of the data import pipeline. This AI-assisted approach delivered a prototype four times faster than the manually written version, while also reducing memory usage to just 1/20th of the original. This significantly expedited the system construction process. Anthropic also emphasized that engineers must make architectural trade-offs, carefully review AI-generated solutions, reject unreasonable AI outputs, clarify system logic, and oversee the overall design. These steps are crucial to ensure that the time saved by AI-generated code is not offset by subsequent delays and rework.