According to a detailed engineering retrospective published by Anthropic Engineering on October 13, 2026, the artificial intelligence research company has fully abandoned Git for its primary repository, operating entirely on a proprietary distributed source control engine named Scion.
The migration was precipitated by an unprecedented explosion in internal code synthesis. As engineering teams deployed autonomous instances of Claude to iterate on research loops, refactor internal tools, and build services, daily merge throughput escalated from approximately 2,500 pull requests in May 2026 to over 30,000 by September. With the repository's main branch history exceeding 100 gigabytes and a compressed working copy surpassing 25 gigabytes, continuous integration systems were executing roughly 25 million code checkouts every 24 hours. Naive Git operations would have generated over 600 petabytes of network transfer daily.
To sustain this volume, Anthropic initially maintained several hundred internal Git mirrors. However, these servers became operationally unviable as merge rates accelerated; repository seeds lagged behind the tip of main, and partial clone operations like diff and blame routinely degraded. The team realized that Git's decentralized model—in which every clone is a standalone, offline-capable database—was fundamentally mismatched with autonomous agent workloads where every worker, sandbox, and CI runner is permanently connected to the network.
Scion replaces the Git commit tree with atomic 'changesets' recorded directly into a write-ahead log. Instead of storing tree pointers and Git packfiles, each changeset contains a list of modified files relative to its first parent and references content-addressed blobs stored in cloud object storage. Global branch references are decoupled entirely from commit creation, existing as compare-and-swap rows in Google Cloud Spanner databases. Because changesets require no cross-server locks or global packfile compression, the server fleet can ingest commits with near-linear horizontal scalability.
For client interactions, Anthropic deployed a three-tier caching topology consisting of a local machine daemon (`scion-daemon`), an in-memory regional cache (`scion-cache`), and authoritative storage handlers (`scion-server`). Rather than negotiating delta compressions, clients query missing content-addressed objects recursively, achieving an average cache hit rate exceeding 99 percent. CI checkouts that previously required minutes now complete in seconds, yielding an immediate 25 percent reduction in total CI compute capacity requirements across the organization.
Remarkably, the entire platform—encompassing Scion, a code review interface named Antique, semantic code search, and directory-level identity and access management—was implemented by a team of seven engineers in five months. The team credited an internal snapshot of Claude Mythos Preview with accelerating development, noting that Claude autonomously rewrote an early handwritten ingestion pipeline to run four times faster with twenty times less memory. What remains unproven is whether other large software enterprises can adapt similar architectures without building custom code review and CI ecosystems from scratch.