NewsAI & DevelopmentDeveloper Tools

Anthropic Scion: Git Broke at 30,000 Merges a Day

Anthropic’s monorepo hit 30,000 merges a day in September — one every three seconds, around the clock — and Git fell over. So the company built a replacement from scratch, shipped it in five months, and called it Scion. It won’t be open-sourced. But the architecture tells you exactly where version control is headed now that agents write more code than humans do.

Git Was Not Designed for This

Between May and September 2026, Anthropic’s daily merge count grew 12x — from roughly 2,500 to more than 30,000. That jump is almost entirely attributed to agents. Claude is now writing more of Anthropic’s internal code than its human engineers are.

Git’s model assumes developers clone a repo, pull occasionally, and push when done. That works fine for a team of a few hundred engineers. It does not work when CI is performing 25 million checkouts a day against a repo with 100 GB of main-branch history and a 25 GB compressed working copy. Each agent spin-up needs fresh environment access. Multiply that across thousands of parallel agents and you get infrastructure costs that compound fast.

What Scion Does Differently

Scion ditches the distributed-clone model entirely. It is an always-online client-server system: clients stay connected, don’t hold full repos locally, and query source control via API. The analogy that nails the distinction: Git is a DAG of commits designed for humans to reason about history. Scion is a write-ahead log designed for machines to merge fast.

The architecture runs three tiers. A scion-daemon runs on every machine and stores objects for that node, starting from a periodically refreshed snapshot. A regional scion-cache sits above it, holding trees and blobs. At the top, the authoritative scion-server stores changeset metadata in Spanner and file content in cloud storage. All three tiers respond to identical RPCs, so clients don’t need to understand the topology — they just ask for data and get it from the nearest cache.

The unit of work in Scion is a changeset, not a commit. A changeset is a write-ahead log entry: it lists which files changed relative to the first parent and points to content-addressed blobs. References are updated with a compare-and-swap on a single Spanner row. Because all data is immutable, every server tier is just a cache with no coordination overhead. That is what makes 30,000 merges a day survivable.

The Results

  • 200x less CPU and memory than the old Git mirror infrastructure
  • Checkout time: minutes → seconds
  • 25% CI CPU capacity freed after the migration
  • Cache hit rate: 99%+ at every tier

CI jobs that used to need a local repository now hit the source control API directly and skip the clone entirely. Seven engineers built this in five months. Anthropic’s engineering post on Scion notes that every engineer on the team would rate Claude as better at writing code than themselves — but that engineers still supplied the design-level architecture thinking that made it work. That is a quietly remarkable statement about where human contribution is concentrating as AI tooling matures.

Anthropic Is Not the Only One Rewriting Source Control

This is where the story gets broader. On November 2, OpenAI disclosed Pando, its in-house Git backend built for engineers and agents working simultaneously. Two days later, OpenAI announced it is acquiring Pierre Computer Company, whose open-source Trees and Diffs libraries already run inside Codex, to bring agent-scale code storage and review into the platform.

Cursor launched Origin, a Git-compatible forge that exposes an API and MCP server so agents can operate it directly — open pull requests, read review state, resolve conflicts, and retry failed builds without a human in the loop. A live demo showed 22.6 commits per second and hundreds of thousands of clones per hour.

The pattern is clear. Git the protocol is not going anywhere. Git the client model — clone, pull, push — is being replaced at the top end of the market. The economic value in source control is migrating to the server.

What to Do Before You Hit the Same Wall

Scion won’t be released publicly. But the architecture signals what to optimize for right now:

  • Use sparse checkouts and shallow clones to reduce what agents pull on startup
  • Push source control operations server-side wherever possible — reduce clone-on-every-job patterns in CI
  • Evaluate Cursor Origin (Git-compatible, paid plans) or Meta’s open-source Sapling for larger monorepos
  • Model your checkout volume per agent now, before it becomes a bottleneck

Agents will eventually break your version control too. Anthropic hit the wall after a 12x growth spike in four months. The question for every team scaling agentic workflows isn’t whether this happens — it’s when, and whether you’ve built the infrastructure to handle it.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News