Anthropic Rebuilt Git (yes really)

Summary
Theo Browne (t3.gg) reviews and breaks down an Anthropic engineering blog post titled “Redesigning source control at Anthropic”, which details how Anthropic built and migrated to an in-house version control system named Scion to replace Git for their monorepo. Theo examines the architectural principles of Scion, why Git’s offline and decentralized model breaks down under large-scale automated agent workloads, and shares his own benchmarks and projects reflecting the shift toward custom tooling.

What is shown

• [00:47] Screen display of “The Missing GitHub Status Page” showing GitHub platform uptime and outages.
• [01:41] Anthropic’s post on X announcing their engineering blog post on replacing Git with an in-house version control system.
• [03:58] The full Anthropic engineering blog post “Redesigning source control at Anthropic” (published October 13, 2026), read and analyzed on screen.
• [05:47] Theo’s GitHub commit history in the pingdotgg/t3code repository, demonstrating merging over 50 PRs in a single 4-hour late-night session.
• [07:19] Anthropic monorepo merge volume chart tracking daily PRs and Antique change requests from January to September 2026.
• [08:15] Website interfaces for Graphite and Cursor.
• [09:40] OpenAI web interface and 404 page for github.com/openai/openai to illustrate monorepo architecture.
• [14:41] Diagrams detailing Scion’s changeset data structures, Spanner metadata tracking, and blob storage layout.
• [21:15] Architectural diagram showing the relationships between clients, scion-daemon, regional scion-cache, authoritative scion-server, Google Cloud Spanner, and Cloud Storage.
• [26:47] Detailed breakdown of Scion checkout architecture and caching layers.
• [28:30] Theo’s side project Lust / lust-lang running in an AI-assisted agent terminal to compile a subset of Rust.
• [30:24] & [33:36] Analysis of CI time overhead in T3 Code using AI agent queries, calculating ~1,000 runner-hours spent waiting on Git setup.
• [32:28] NPM package page for tsc-rs showing 42,000+ weekly downloads.
• [37:06] Anthropic benchmark bar charts comparing Scion vs. Git execution times across developer and agent environments.
• [39:23] Claude status dashboard displaying high operational uptime across Claude services.

Claims & numbers

• The presenter notes Anthropic grew from merging ~2,500 code changes per day in May 2026 to over 30,000 per day in September 2026.
• Anthropic migrated its monorepo from Git to Scion in five months with a core team of seven engineers.
• Anthropic’s monorepo main branch history is nearly 100 GB, with compressed working copies exceeding 25 GB.
• Anthropic executes approximately 25 million code checkouts per day for CI alone; fetching naively over Git would exceed 600 PB of traffic daily.
• Scion change ingestion was prototyped by human engineers, but an early internal preview of Claude Mythos rewrote the Git-to-Scion pipeline, reducing memory consumption by 20x and increasing execution speed by 4x without changing implicit invariants.
• Scion maintains a cache hit rate of over 99%, and its infrastructure footprint is 200x smaller than Anthropic’s previous Git mirroring setup in terms of CPU and memory.
• In CI and cluster jobs, Scion checkouts execute two orders of magnitude faster (seconds rather than minutes), yielding an immediate 25% CPU capacity win for CI workloads alone.
• In developer environments, Scion commands outperform Git: pulling latest is 25x faster (3.5s vs 87s), checkouts are 14x faster (1.1s vs 15s), rebases are 4.8x faster (2.7s vs 13s), status is 207x faster (19ms vs 3.9s), diffs are 6.1x faster (67ms vs 407ms), and worktree creation is 84x faster (5.1s vs 7.1 min).
• Theo calculates that t3code spent ~1,000 runner-hours (around 42 days of cumulative runner time across ~465,000 checkouts) costing roughly $350–$600 solely waiting for Git operations.

Notable quotes

• [00:42] “Git is not the right primitive for the way that things are moving right now.” — Theo Browne
• [28:24] “Now we can build software for our specific needs.” — Theo Browne
• [44:08] “Claude is now better at dealing with merge conflicts than I am.” — Theo Browne

Assessment
This is an authentic commentary and technical review video by an independent software engineer and streamer analyzing an official Anthropic engineering blog post. The presenter demonstrates his own genuine terminal logs, project stats, and CI data while reading directly through the published technical documentation without deceptive staging or unverified claims.

Described by gemini-3.8-flash on 2026-10-16 from the video’s audio and frames.