Anthropic rebuilt its source-control stack because development velocity exploded: daily merged changes rose from about 2,500 in May to over 30,000 by September, with a peak above 30,000 on September 9. The main branch history approaches 100 GB and a compressed working copy exceeds 25 GB, while CI alone performs on the order of 25 million checkouts per day - traffic that would amount to hundreds of petabytes if served by naive Git. To meet those scale and performance constraints, a seven‑person team created Scion, a custom system that leans toward a centralized, server-backed model and adopts the concept of derived data (building on ideas similar to Google’s Piper and Meta’s Sapling/Mononoke). They also introduced a new review workflow called “antique” which rapidly replaced pull requests as the dominant change type during the summer.
The redesign traded some traditional Git features for operational efficiency: migration from Git completed in five months, and Scion intentionally omits offline support because the largest consumer (their Claude models) is always online. Anthropic makes extensive use of adversarial review - having models find flaws - which speeds discovery but can be myopic and requires experienced engineers to avoid over‑engineering. Model release timing (notably the Fable/Mythos class) correlates with the surge in commits, suggesting product shifts drove much of the increased coding velocity.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.