When coding agents multiply merge volume more than tenfold, source control, builds and CI become the bottleneck. Anthropic shows what broke in git at monorepo scale and how they rebuilt it, a preview of problems other teams will hit.
Key takeaways · AI-distilled
Anthropic's main-branch history is almost 100 GB with a 25 GB+ compressed working copy, and CI does about 25 million checkouts a day, which the post says would be over 600 PB of git traffic daily if fetched naively.
A Scion changeset stores only files changed versus its first parent plus content-addressed blob pointers; blame, path history and commit-graph skip lists are derived data computed recursively from parents, eagerly or on demand.
The client requests only the trees and blobs it wants rather than telling the server what it has, so any cache tier can answer; the post reports 99%+ hit rates per tier and a footprint 200x smaller than the old git mirrors.
A team of seven built Scion and the Antique review tool and migrated in five months; engineers often had to tell Claude no, as when systems linearly scanned the commit log or adversarial review pushed toward excess validation.
Anthropic dropped a global merge-queue lock for post-merge and pre-deploy testing, accepting conflicts that Claude fixes, and says Scion is no general git replacement; it suggests git forges could offer server-side RPCs instead.
