Urgent.News

What's breaking now, across thousands of outlets.

AI

Redesigning Source Control at Anthropic

Anthropic has unveiled a new approach to source control, announcing the development of their own custom system called Scion. The decision to create Scion arose from a dramatic surge in code changes at Anthropic, which went from around 2,500 merged per day in May 2026 to over 30,000 per day by September. This growth was attributed, in part, to the introduction of larger models like Fable and Mythos.

Scion functions as a reinvention of centralized source control systems, similar to Subversion. The history of the main branch alone stands at nearly 100 GB, with compressed working copies reaching over 25 GB. To manage the vast amount of data, Scion handles about 25 million code checkouts daily for CI, excluding developer environments, agent sandboxes, and research jobs. This amount of data would generate over 600 petabytes of Git traffic daily if fetched naively.

The new system was developed by a team of seven engineers, and the transition from Git was accomplished in five months. When inquired about the absence of offline support, a Scion engineer named Nick G explained that it was unnecessary since Claude, Anthropic's primary user of the system, operates exclusively online.

Anthropic also employs adversarial review extensively, often pitting models against each other to identify flaws. This practice has proven effective; however, engineers must intervene to prevent over-engineering in response to speculative concerns. Simon Willison, a blogger, shared his thoughts on the subject, noting that while adversarial review is highly beneficial, it can sometimes lead to a narrow focus on hypothetical issues.

Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Read the original at simonwillison.net →

More in AI

More from Friday 16 October →