{"slug": "anthropic-scion-when-claude-wrote-30000-merges-a-day", "title": "Anthropic Scion: When Claude Wrote 30,000 Merges a Day", "summary": "Anthropic built an in-house version control system called Scion after its monorepo's daily merges grew from 2,500 in May 2026 to more than 30,000 by September 2026, a 12x increase in four months. A seven-engineer team built and migrated to Scion in five months, alongside a code review surface called Antique, a dedicated code search system and a source control IAM layer, replacing Git, GitHub and third-party review tooling; Anthropic says the server-side changeset design cut infrastructure footprint 200x versus its Git mirror fleet, dropped CI checkout times from minutes to seconds and added 25% CPU capacity across CI. Anthropic is not releasing Scion, saying it only makes economic sense for organizations with massive monorepos, and expects heavy agent users to hit this class of problem within six to twelve months.", "body_md": "In May 2026, Anthropic’s monorepo was merging 2,500 code changes a day. By September — four months later — that number crossed 30,000. Git didn’t fail dramatically. It just became untenable: 100 GB of compressed history, 25 million CI checkouts per day, a working copy over 25 GB. The infrastructure required to serve all of that via standard Git mirroring was unsustainable. So Anthropic built something new.\n\nThe result is **Scion**, an in-house version control system designed for a codebase where AI agents do most of the committing. A seven-engineer team built and migrated to it in five months. They also shipped **Antique** (a code review surface), a dedicated code search system, and a source control IAM layer — replacing Git, GitHub, and third-party review tooling in the process.\n\n## Why Git Broke\n\nGit was designed for distributed human developers who work offline, clone full repositories, and push at human pace. Anthropic’s agents don’t operate that way. They’re always online, they don’t need full clones, and they generate commits at a rate no human team approaches. The daily merge volume grew 12x in four months. Serving 25 million daily CI checkouts via naive full-fetch would have required roughly 600 petabytes of daily data transfer. No off-the-shelf Git setup survives that math.\n\nBeyond bandwidth, Git’s client-side computation model hits a wall at agent scale. Diff, blame, and merge operations that make sense to run locally for one developer become absurd when thousands of agents are doing them simultaneously. Git was never designed for this.\n\n## What Scion Does Differently\n\nScion flips the model: the server does the work, clients cache only what they need. Its unit of change is a **changeset** stored in a write-ahead log — a delta against the parent, not a full-tree snapshot like a Git commit. Every commit, merge, and blame operation happens server-side. Clients maintain a bounded local cache; they don’t clone, they don’t push in the conventional sense.\n\nThe checkout architecture runs three tiers — authoritative server (Spanner-backed), regional cache, per-node daemon — achieving cache hit rates above 99% at every level. All objects are content-addressed: never change once written, cacheable by any server without coordination. No packfiles to maintain, no global merge queue locks. The result: a 200x reduction in infrastructure footprint versus the Git mirror fleet, CI checkout times dropping from minutes to seconds, and a 25% CPU capacity gain across CI.\n\n## The Catch\n\nAnthropic is not releasing Scion. They’re clear that it’s not a general-purpose Git replacement and only makes economic sense for organizations with massive monorepos — tens of thousands of engineers or swarms of coding agents. Building Scion also required building a new code review tool, new search, and a new IAM layer. It’s a full platform replacement.\n\nThe skeptical take on this is fair. Anthropic’s [engineering post](https://www.anthropic.com/engineering/redesigning-source-control) generated nearly 1,000 Hacker News comments, and plenty of them noted that 30,000 merges a day is an Anthropic-scale problem funded by Anthropic-scale token spend. Most teams are nowhere near this. The question is where they’re headed.\n\n## Why This Still Matters for Your Team\n\nThe story isn’t Scion itself — you’re not going to use it. The story is the trajectory. Teams running a few dozen agents against a shared repo are on a smaller version of the same curve. [GitHub infrastructure has been straining under AI agent load](https://www.theregister.com/devops/2026/05/15/git-is-unprepared-for-the-ai-coding-tsunami/5241480) throughout 2026, with outages becoming routine. Anthropic expects heavy agent users to hit this class of problem within six to twelve months.\n\nCursor saw it coming. Their [Origin platform](https://www.infoq.com/news/2026/08/cursor-origin-alternative-github/), now in early beta, keeps Git as the underlying protocol but builds an agent-native forge around it — stacked PRs, agent-integrated merge queues, and editor plus repo in the same surface. It’s a more practical path than what Anthropic built, and one most teams can actually adopt.\n\nThe read from the Anthropic post: the bottleneck in an AI-heavy codebase has shifted from writing code to merging it. Review automation and merge queue intelligence are the new competitive moat. If you’re scaling agent usage seriously, watch what Git platforms add to their server-side APIs over the next year — that’s where the pressure Anthropic exposed will get resolved for the rest of the industry, not through everyone building their own Scion.\n\nPractical steps now: audit how many daily merges your agents are generating, track whether CI queue times are trending up, and check [Simon Willison’s notes on the original post](https://simonwillison.net/2026/Oct/13/redesigning-source-control-at-anthropic/) for the clearest independent summary of what Anthropic actually built.", "url": "https://wpnews.pro/news/anthropic-scion-when-claude-wrote-30000-merges-a-day", "canonical_source": "https://byteiota.com/anthropic-scion-git-ai-scale/", "published_at": "2026-10-17 09:08:34+00:00", "updated_at": "2026-10-17 10:03:55.167757+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "ai-tools", "artificial-intelligence"], "entities": ["Anthropic", "Scion", "Antique", "Git", "GitHub", "Cursor", "Origin", "Hacker News"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anthropic-scion-when-claude-wrote-30000-merges-a-day", "markdown": "https://wpnews.pro/news/anthropic-scion-when-claude-wrote-30000-merges-a-day.md", "text": "https://wpnews.pro/news/anthropic-scion-when-claude-wrote-30000-merges-a-day.txt", "jsonld": "https://wpnews.pro/news/anthropic-scion-when-claude-wrote-30000-merges-a-day.jsonld"}}