Technology
Anthropic built a new source-control system for software written at agent scale
Anthropic says it replaced Git for its main monorepo after daily merged changes rose from about 2,500 in May to more than 30,000 in September. Its Scion system assumes always-online agents, server-side operations, sparse or virtualized checkouts and horizontally scalable caches. The company reports a 200-fold smaller infrastructure footprint, checkout speeds two orders of magnitude faster and a 25% CI CPU reduction. Anthropic stresses that Scion is bespoke and unsuitable for ordinary repositories, but expects agent-driven workloads to create similar pressures elsewhere.
Editorial summary · Publisher article
What matters
- Scion stores incremental changesets and computes history, blame and other derived data through server-side services.
- The system avoids fully materialized repository copies and uses tiered caches with reported hit rates above 99%.
- A seven-person team built Scion and related tooling and migrated from Git in five months.
- Engineers still set system constraints, review intent and steer agents away from over-engineering or invalid designs.
Why Git stopped fitting the workload
Anthropic says its main-branch history approached 100 GB, compressed working copies exceeded 25 GB and CI performed roughly 25 million checkouts per day. Git mirrors and partial-clone workarounds became harder to operate as merge rates increased, especially because most clients were online agents rather than offline human workstations.
A server-centered architecture
Scion moves commits, rebases, merges and repository queries to scalable services. Clients keep bounded caches and request only the trees and blobs they need, while server-side derived data supports operations such as blame, path history and ancestry queries without requiring every job to maintain a local repository.
Measured gains and broader limits
Anthropic reports that Scion’s infrastructure is 200 times smaller in CPU and memory terms than its previous Git-mirroring setup, with checkouts taking seconds rather than minutes and CI using 25% less CPU. The company says the design is intended for enormous monorepos with prolific agents, not general Git users, and acknowledges migration costs for organizations that would need matching review, search and CI systems.
Context & caveats
- The performance figures are Anthropic’s own operational measurements, and the system is described by its builders rather than independently audited.
- The report presents an early, highly specialized design; its applicability to ordinary engineering teams remains limited.
Why this made the edition
Anthropic provides unusually detailed primary evidence about rebuilding source control for agent-scale software development, including architecture, measured performance and explicit limitations.
Source material