Redesigning source control at Anthropic
by Anthropic
How Anthropic replaced git for its monorepo once coding agents pushed merges past 30,000 a day.
Overview
Redesigning source control at Anthropic, published October 13, 2026 by Nick G. and Anthropic's source control team (updated October 15 with more detail on prior art), describes how a team of seven built Scion, an in-house version control system, in five months and moved Anthropic's monorepo off git. The trigger was agent volume: daily merges rose from about 2,500 in May 2026 to over 30,000 in September, main-branch history neared 100 GB and CI ran about 25 million checkouts a day. The post has five parts. Growing our monorepo covers the git workarounds that stopped scaling (shallow and sparse clones, promisor remotes, several hundred mirrors). Principles for agent-scale source control lists the design rules: always-online clients, server-side history, blame, merges and rebases, source control as an API, no full repo copy anywhere, cheap lock-free server-side commits and optional sparse working copies. How Scion works explains changesets stored as write-ahead log entries with content-addressed blobs, lazily or eagerly computed derived data, ref updates by compare-and-swap on a Spanner row and a three-tier cache topology, which Anthropic reports cut infrastructure CPU and memory about 200x versus git mirroring. What we learned covers agent-written code, human review of intent, semantic approval rules and dropping a global merge queue. The last part says Scion will not be released. The design credits Google's Piper and Meta's Sapling and Mononoke; a former Mononoke engineer argued on sunshowers.io that the original credit was inadequate, and Anthropic later added attribution.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Guide
- Cost
- Free
- Duration
- ~15 min read
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Why git's offline, decentralized design strains when always-online agents and CI jobs are the main clients
- ✓Which git workarounds Anthropic tried first, including shallow clones, sparse checkouts, promisor remotes and mirrors
- ✓Six design principles for agent-scale source control, from server-side merges to optional sparse working copies
- ✓How changesets, content-addressed blobs and lazily computed derived data replace packfiles and full clones
- ✓How a three-tier daemon, regional cache and server topology keeps cache hit rates above 99 percent
- ✓Why a global-lock merge queue fails at agent scale and what post-merge testing replaces it with
- ✓How semantic approval rules separate trivial agent changes from sensitive ones in code review
- ✓Where human engineers still add value when Claude writes most of the code: architecture, tradeoffs and saying no
Highlights
- •Rare hard numbers on agent-driven engineering load: 2,500 to 30,000+ merges a day between May and September 2026, ~25 million CI checkouts a day
- •Reports concrete results (about 200x less infrastructure CPU and memory than git mirroring, checkouts in seconds, 25% more CI CPU capacity) rather than vague productivity claims
- •One of the most discussed engineering posts of the month: 899 points and 1,013 comments on Hacker News as of 2026-10-25
- •Read it with the prior-art debate: a former Meta Mononoke engineer's critique on sunshowers.io prompted Anthropic to add attribution, which helps you judge what is new
- •Scion itself is not released, so this is an architecture case study, not a tool you can install
Who It’s For
Best For
- ✓Platform and developer-productivity engineers preparing CI and source control for coding-agent traffic
- ✓Engineering leaders deciding how review and merge workflows should change when agents write most code
- ✓Infrastructure engineers comparing monorepo designs such as Piper, Sapling and Mononoke
Prerequisites
- •Working knowledge of git internals: commits, refs, packfiles, clones and sparse checkouts
- •Familiarity with CI pipelines, monorepos and distributed caching
- •Some experience with coding agents such as Claude Code helps you follow the workflow sections
FAQ
What is Redesigning source control at Anthropic?
An Anthropic engineering write-up for platform and developer-infrastructure engineers on what breaks when coding agents become the main users of source control. It explains Scion, the in-house version control system that replaced git for Anthropic's monorepo, and the design principles you can borrow for agent-scale CI and code review.
Is Redesigning source control at Anthropic free?
Redesigning source control at Anthropic is free to access.
What level is Redesigning source control at Anthropic for?
Redesigning source control at Anthropic is aimed at a advanced audience. Recommended background: Working knowledge of git internals: commits, refs, packfiles, clones and sparse checkouts, Familiarity with CI pipelines, monorepos and distributed caching, Some experience with coding agents such as Claude Code helps you follow the workflow sections.
How long does Redesigning source control at Anthropic take?
Expect roughly ~15 min read. Most learners work through it at their own pace.
What will I learn from Redesigning source control at Anthropic?
You'll learn: Why git's offline, decentralized design strains when always-online agents and CI jobs are the main clients; Which git workarounds Anthropic tried first, including shallow clones, sparse checkouts, promisor remotes and mirrors; Six design principles for agent-scale source control, from server-side merges to optional sparse working copies; How changesets, content-addressed blobs and lazily computed derived data replace packfiles and full clones; How a three-tier daemon, regional cache and server topology keeps cache hit rates above 99 percent; Why a global-lock merge queue fails at agent scale and what post-merge testing replaces it with; How semantic approval rules separate trivial agent changes from sensitive ones in code review; Where human engineers still add value when Claude writes most of the code: architecture, tradeoffs and saying no.
Topics
Sources
This page was written from 5 sources, 3 on domains other than anthropic.com.