Anthropic published an engineering post on October 13 titled "Redesigning Source Control," describing how a seven-person team spent five months migrating the company's entire monorepo from Git to a purpose-built version control system called Scion. During the migration window, daily code merges grew from about 2,500 in May to more than 30,000 by September—a more than tenfold increase in four months.
Why the bottleneck moved
Anthropic's core claim is blunt: writing code is no longer the bottleneck. The bottleneck has shifted to developer infrastructure—source control, builds, and continuous integration—that was designed for human-scale usage.
The repository's scale supports this: main-branch history is close to 100 GB, compressed working copies exceed 25 GB, and CI alone performs about 25 million checkouts per day. Naive Git usage would generate over 600 PB of daily traffic (a 16 TB drive filled 37,500 times over). Existing mitigations—shallow clones, sparse checkouts, hundreds of internal Git mirrors at peak, and blob-storage snapshots—bought time but degraded as merge rates climbed, breaking tools like diff and blame.
The key insight is a change in usage, not code volume. Git assumes decentralized, offline-capable clones, each a full repository. Anthropic's clients are almost entirely agents and automated jobs that are online by definition—offline capability became pure overhead. The redesign makes clients always-online, with servers doing the heavy lifting and clients keeping only a bounded cache.
Scion's design
Scion borrows from Google's Piper and Meta's Sapling, following four principles:
- Clients are always online
- Clients and servers are asymmetric: servers answer history queries, blame, commit creation, and merges; clients maintain only a bounded local object cache
- All source control capabilities are exposed via API, so jobs can query without local repositories—enabling automated bisection and rollback pipelines at scale
- The system scales horizontally
- CPU and memory requirements reduced to 1/200
- Checkout speeds for CI and research jobs improved by two orders of magnitude, from minutes to seconds
- 25% CPU capacity headroom from CI savings alone
Scion replaces Git's commit with a changeset: a record in a write-ahead log containing the list of files changed relative to the first parent, plus pointers to compressed content in blob storage. Critically, changesets are created server-side and immediately available everywhere—Git commits are local objects with explicit pushes. This difference underpins all the performance numbers.
A three-layer cache structure—node-level scion-daemon, regional scion-cache, and authoritative scion-server—responds to the same RPC set with strict layering to isolate failure domains. Anthropic reports cache hit rates above 99% at every layer.
Measured results versus the prior Git mirror infrastructure:
Claude in the loop
The migration itself involved Claude. An early internal snapshot of Claude Mythos Preview rewrote the previously hand-written Git-to-Scion import pipeline, producing—without further instruction—a version using 20x less memory and running 4x faster, while preserving all implicit invariants around import ordering and data structures.
Anthropic is candid about the human role: engineers now rate Claude a better code writer than themselves, but their work has changed. They still review every change request—reviewing intent rather than lines—and much of the job is telling Claude "no," for example when agents violate core system constraints or over-engineer solutions.
Not for everyone
Anthropic explicitly states Scion is not a general-purpose Git replacement. It targets very large monorepos with swarms of agents and depends on custom code review, search, and CI. It will not be released publicly. The post instead shares design lessons for platforms like GitHub and GitLab—richer server-side RPCs, better combined sparse/shallow features—without abandoning Git's model, and acknowledges the Git community's nearly two decades of scaling work.
For teams far from 30,000 daily merges, Git remains the standard. But Anthropic predicts similar problems will appear in more organizations within six to twelve months as AI coding tools spread.
What to watch
Anthropic's stated premise is that "the latent demand for merged code is effectively unlimited": removing every bottleneck immediately triggers more code output. This creates an uncomfortable loop—faster code, infrastructure bottlenecks, rebuilt infrastructure, more code, next bottleneck. A September retrospective on CI showed the same pattern: task volume grew 25x in six months, three stopgap fixes lasted 70 days, 29 days, and under a day respectively, before one engineer rebuilt the test-selection service in three weeks.
Two things to watch: first, whether other organizations announce similar migrations in the next 6–12 months. Second, the regional cache layer—99% hit rates mean ~1% of requests hit the authoritative server, and the tail latency of that path during a regional failure is unaddressed in the post. Finally, the team dropped the global merge-queue lock in favor of post-merge and pre-deploy testing, accepting concurrent conflicts on the assumption that agents can fix them quickly—a premise that depends on real agent bug-fixing capability.