With the rapid advance of AI coding capabilities, Anthropic's internal development efficiency has seen an unprecedented explosive surge. According to officially disclosed data, in the 5th month of this year (May) its average daily number of code merges was approximately 2500, while by the 9th month (September) that figure had soared to more than 30,000. This explosive rate of code growth has left the original Git version control system and other development tools struggling to keep up.
Facing a continuous integration service that was overwhelmed, the team had previously made three temporary fixes, but each one lasted a shorter time, until they finally had to completely solve the problem through a rewrite. To thoroughly break through the performance bottleneck of source code management, Anthropic's 7-person R&D team spent 5 months building Scion, a brand-new version control system purpose-built for ultra-large-scale codebases, and successfully fully migrated the company's massive monolithic codebase from Git to it.
In terms of architecture, Scion adopts an innovative model in which the server handles core operations while clients fetch data on demand, complemented by an efficient layered caching mechanism. This optimization reduced the scale of its source code management-related infrastructure to just 1/200 of the previous Git mirror infrastructure. Notably, during system development, the team used Claude to rewrite the data import pipeline prototype, which ran 4 times as fast as the handwritten version while consuming only 1/20 of the original memory, greatly accelerating the construction timeline of the entire system.
While enjoying the tenfold productivity gains that AI brings, Anthropic has also distilled a set of valuable hands-on lessons. The team points out that engineers must be prepared to make architectural trade-offs, rigorously review AI-generated solutions, dare to say "no" to unreasonable AI output, and at the same time clarify system logic and keep overall design under control—only then can the time saved by AI-generated code truly avoid being consumed by later waiting and frequent rework.