Redesigning source control at Anthropic
Anthropic has built “Scion”, a custom centralized source control system to replace git for its rapidly growing monorepo and fleets of coding agents, which now generate tens of thousands of merges and tens of millions of code checkouts per day. Commenters compare the design to Perforce or Subversion and debate whether the real problem is git’s scalability, Anthropic’s monorepo and workflow choices, or the sheer volume of AI‑generated churn. The thread raises broader questions about agent‑driven development: how to review intent rather than every line of code, whether such velocity is sustainable or wise, and why Anthropic is not open‑sourcing tooling that was trained on so much external code.
Why Anthropic built Scion
- Git started failing at Anthropic’s scale: ~100 GB main-branch history, ~25 GB working copy, ~25M checkouts/day for CI and agents.
- Naively cloning per job would imply ~600 PB/day of traffic; even with caching and repacking, people report git repo maintenance becoming brittle.
- Monorepo is intentional: enables atomic cross-cutting changes, consistent tooling, and easy agent access, but amplifies scaling problems.
Design and comparisons
- Scion is centralized and always-online: all commits/merges/rebases happen on the server; clients fetch only needed blobs on demand.
- History and file contents are immutable, content-addressed objects stored in a log + object store, with derived data computed asynchronously.
- Many commenters say this “sounds like” Subversion, Perforce, ClearCase, or Google’s Piper / Meta’s Sapling, i.e. central server + virtualized working trees.
- Others point out similar ideas in EdenFS, artifact-based filesystems, and jj-based tooling; some see it as an inevitable pattern for huge monorepos.
Scale, agents, and workflow changes
- Merges grew from ~2.5k/day to >30k/day, largely driven by agents. Some see this as “gradient descent with code” or brute-force exploration.
- CI/merge-queue semantics shift: you can’t globally lock the repo, so more is done post-merge and pre-deploy; concurrent semantic conflicts are expected and fixed by agents.
- Human engineers’ role is described as choosing problems, making architectural tradeoffs, and frequently telling agents “no,” then reviewing intent rather than every line.
Quality, safety, and maintainability concerns
- Many are uneasy with 30k merges/day and high churn: fear of spaghetti code, jank, and “AI slop” accumulating faster than it can be understood.
- Several note that even today they must closely review AI-generated code to avoid subtle bugs; they doubt Anthropic can truly oversee everything at this velocity.
- Centralized, online-only VCS raises worries about outages, security, and loss of offline workflows; others reply that large orgs already live in always-online, API-driven setups.
Open-sourcing, generality, and broader implications
- Anthropic says Scion is too bespoke and too divergent from git to open-source usefully; many readers are frustrated by a detailed blog post without code.
- Some argue most organizations should stick to git or buy Perforce-like systems; only frontier labs with huge agent fleets and token budgets need Scion-style tooling.
- Broader debate: LLMs make it cheap to reinvent infrastructure (VCS, CI, deploy, dashboards), but it’s unclear how much of that yields lasting value vs churn and tech debt.