Anthropic Rebuilt Git (yes really)

Summary

Theo examines Anthropic’s in-house version-control system, Scion, built in response to a sharp increase in AI-assisted code production. Anthropic’s reported merge rate rose from about 2,500 changes per day in May to more than 30,000 in September. The company says developer infrastructure designed around human workflows—especially source control, CI, and code review—had become a bottleneck.

Why Anthropic Moved Beyond Git

Anthropic’s monorepo had grown difficult to serve using conventional Git infrastructure: its main-branch history was reported to be nearly 100 GB, with a compressed working copy over 25 GB. CI alone performed roughly 25 million checkouts per day. Git’s model, in which each clone holds a complete, independently capable repository, worked well for distributed, often-offline human development but created substantial storage, traffic, and maintenance costs at Anthropic’s scale.

Theo agrees that Git and GitHub can be friction points for highly parallel agent workflows, while noting that companies with unusual needs have long customized or replaced Git. He cites Meta’s Mercurial-based tooling and other source-control products as precedent. He also stresses that Anthropic’s solution is specialized—not a general-purpose replacement for Git.

How Scion Works

Scion was developed over five months alongside Antique, Anthropic’s code-review interface, plus code search and other source-control tools. A team of seven reportedly built Scion and Antique and migrated related systems during that period.

Its design differs from Git in several key ways:

• Online, server-centered operations: Since most clients are agents and CI jobs that are online, Scion’s server handles operations such as creating commits, merging, and rebasing. Clients keep a bounded local cache and working copy rather than a full repository.
• Change sets and immutable storage: The unit of work is a change set, recording changed files and content-addressed pointers to their contents in blob storage. Change sets are created server-side and are available across the system without a push step.
• Derived data for queries: Scion builds additional structures for tasks such as blame, path history, and commit-graph queries. These can be computed eagerly or deferred, and changed without exposing their internals to clients.
• Sparse and virtual checkouts: Clients can request only the directories, trees, or files they need. Some files can be materialized on demand, and many CI or source-control queries can use APIs without checking out a working copy.
• Tiered, horizontally scalable caching: Regional caches and job-oriented daemons serve frequently requested data; authoritative servers retrieve it from storage when needed. The tiers use the same RPC interface but are kept separate to limit the impact of failures.
• Less shared mutable state: Scion minimizes global refs and locking. Its immutable, content-addressed data can be cached broadly without extensive coordination.

Anthropic reports that each cache tier achieves a hit rate above 99% and that Scion’s infrastructure uses about 200 times less CPU and memory than its previous Git-mirroring setup. CI and research jobs can check out code in seconds rather than minutes; the article also reports a substantial reduction in CI resource use.

Workflow Changes and Engineering Lessons

Anthropic says it is rethinking workflows built around slower, human-scale development. Rather than rely on a global merge queue to serialize changes, it has invested more in automated testing after merges and before deployment. Theo argues this is a sensible trade-off at high throughput: teams should make it easier to detect, contain, and revert bad changes, not only add friction to prevent them. Anthropic notes that concurrent changes can conflict, but says Claude can often resolve conflicts quickly.

The team uses Claude extensively to build and improve the system. Anthropic gives an example in which Claude rewrote an ingestion pipeline to use 20 times less memory and run four times faster. It also reports that Claude found algorithmic optimizations and helped generate extensive simulations, fuzzing, and tests.

Human engineers remain important, but their work shifts toward system design, trade-offs, and directing agents. The team reviews the intent of changes rather than necessarily reading every line. Anthropic also warns that agents can violate system constraints, overengineer solutions, or react too defensively to review feedback; engineers need to steer them.

For operations, Anthropic describes using Claude-based “super fans” in Slack to watch component telemetry and logs for subtle issues, alongside conventional alerts.

Performance and Wider Implications

Anthropic reports large improvements in common operations, including 25× faster pulls, 14× faster checkouts, 5× faster rebases, 207× faster status checks in developer environments, and 84× faster worktree creation.

Theo also illustrates the cost of Git checkout delays with his own project: he estimates nearly 1,000 CI hours spent waiting for checkouts, costing roughly $350–$600 in CI compute alone.

Theo’s broader takeaway is that AI lowers the cost of building bespoke tools, making it more practical for companies to replace infrastructure that does not fit their needs. He cautions that Scion’s architecture is aimed at enormous monorepos and agent-heavy development, and that adopting it requires replacing or integrating other parts of the development lifecycle. Anthropic does not present it as a broadly suitable Git replacement; instead, it suggests that Git hosting platforms could adopt some ideas, such as server-side APIs, better sparse-data handling, and alternative server architectures.

Reviews, Guides, or Tutorials

No hands-on tutorial or product review is presented. The video is a technical deep dive and analysis of Anthropic’s published Scion design and the workflow changes around it.

Main Speakers and Sources

• Theo (T3.gg): Narrates the video and offers commentary, examples, and opinions.
• Anthropic Engineering: The source article, “Redesigning source control at Anthropic,” provides the Scion architecture, metrics, and lessons.
• Infinite Red: Briefly appears as the video’s sponsor, in a segment about mobile development consulting.