Back to list

AI Tech36氪

Seven People, Five Months: Anthropic Built a Git Replacement

36氪1 min read
Read original
Seven People, Five Months: Anthropic Built a Git Replacement

AI Summary

Anthropic built Scion, an internal source control system, along with the code review tool Antique, in five months with seven engineers and heavy use of Claude, and migrated its entire monorepo from Git. The move was driven by AI-generated code volume that outgrew Git-based infrastructure, with daily merges rising more than twelvefold in four months. The case suggests other companies may soon face similar bottlenecks.

AI Insights

Anthropic's replacement of Git with Scion, built by just seven engineers in five months using Claude, signals a shift where AI-driven code output overwhelms human-era developer infrastructure. The key implication is that version control and CI tooling designed for distributed human workflows may become bottlenecks, pushing the industry toward always-online, server-centric, API-first source control architectures. Companies adopting AI coding agents should expect similar scaling pressures and begin evaluating their dev infrastructure now.

Analysis written by AI from the original article. Check it against the source.

Translated text

Git, the version control tool that has dominated programming for 20 years, has been replaced by a new system built internally at Anthropic. Seven engineers and Claude, which was deeply involved in development, built the internal source control system Scion, the code review tool Antique, and other related systems in five months, and migrated the entire monorepo from Git to Scion. Anthropic's daily code merges rose from about 2,500 in May to over 30,000 in September, more than a twelvefold increase in four months, and the company expects this growth to continue. Git's design strained under this load: the main branch history alone approaches 100GB, and CI systems perform about 25 million checkouts per day. Despite running hundreds of internal Git mirror servers, Anthropic found these workarounds increasingly hard to maintain. Recognizing that its clients were almost all always-online agents and automated tasks, Anthropic designed Scion around a constantly available server. In Scion, clients fetch only the data they need, the server handles most operations such as history queries and merges, and repository data is distributed so each component can scale independently. The case suggests other companies may soon face similar bottlenecks as AI-driven coding spreads.

Related News

3