A team of 7 relied on Claude to rebuild a new source code system in just 5 months. With the rapid advancement of AI programming capabilities, Anthropic’s internal development efficiency has experienced unprecedented explosive growth.

A team of 7 relied on Claude to rebuild a new source code system in just 5 months. With the rapid advancement of AI programming capabilities, Anthropic’s internal development efficiency has experienced unprecedented explosive growth. According to officially disclosed data, its average daily code merge volume was approximately 2,500 times in May this year, and by September this number had soared to more than 30,000 times. This explosive code growth rate makes the original Git version control system and other development tools begin to be overwhelmed. Faced with the situation where the continuous integration service was overwhelmed, the team had made temporary remedies three times before, but the duration each time was getting shorter and shorter. In the end, the problem could only be solved completely through reconstruction. In order to completely break the performance bottleneck of source code management, Anthropic's 7-person R&D team spent 5 months to build a new version control system specifically for ultra-large code bases - Scion, and successfully migrated the company's huge single code base from Git. In terms of architectural design, Scion adopts an innovative model in which the server handles core operations and the client obtains data on demand, and is equipped with an efficient hierarchical caching mechanism. This optimization reduces the size of its source code management-related infrastructure to only 1/200 of the previous Git mirror infrastructure. It is worth mentioning that during the system development process, the team used Claude to rewrite the data import pipeline prototype. Its running speed reached 4 times that of the handwritten version, while the memory usage was reduced to 1/20 of the original version, which greatly accelerated the construction cycle of the entire system. While enjoying the tenfold productivity improvement brought by AI, Anthropic has also summed up a set of valuable practical experience. The team pointed out that engineers must make choices at the architectural level, strictly review AI-generated solutions, and have the courage to say "no" to unreasonable AI output. At the same time, they must clarify the system logic and control the overall design. Only in this way can the time saved by AI-generated code not be consumed in later waiting and frequent rework. via AI News (author: AI Base)