Artificial intelligence news.
Straight from the original announcements.

Anthropic swaps Git for Scion, its own source control system, on its monorepo

October 19, 2026, 18:51Source: Arize AI
Anthropic swaps Git for Scion, its own source control system, on its monorepo
Image: Arize AI

Anthropic has replaced Git on its large monorepo with Scion, a new server-side source control system that a small team and Claude built and migrated to in about five months. According to an analysis by Arize AI, the move was driven by merge volumes that Git could no longer handle, though most teams are unlikely to face the same limit.

Key points
  • Anthropic replaced Git on its monorepo with Scion, built and migrated to in about five months by a small team and Claude.
  • Merges rose from 2,500 a day in May to over 30,000 in September, and Git could not keep up.
  • According to Anthropic, Scion uses 200x less CPU and memory than the Git mirror fleet it replaced.
  • Scion is not open source, and Anthropic says it is only needed for a monorepo as big as its own.
  • Perplexity and the GitSwarm research system coordinate many agents using ordinary Git.

Why Git stopped scaling for Anthropic

Anthropic's stated reason for the switch was that Git was not scaling. Arize reports that the monorepo went from 2,500 merges a day in May to over 30,000 in September.

Checkouts were a second problem. The main branch's history was almost 100GB, and a compressed working copy was over 25GB. Anthropic ran roughly 25 million CI checkouts a day, which would have produced more than 600 petabytes of daily Git traffic if every checkout fetched the full working copy. The company tried shallow clones, sparse checkouts and hundreds of mirrors of its Git servers, but none of these kept up.

Git's design explains the friction. Every clone copies the whole repository and its history, which suits human developers who may be offline and want a fast local experience. Scion, by contrast, was built for agents working at a very different speed, and for a monorepo of Anthropic's size.

How Scion differs from Git

Scion moves most of the work to the server. According to the Arize post, these are its main design choices:

  • Server-side answers: questions such as who last changed a line, or what the diff between two commits is, are answered through an API call, so the client does not need the tree.
  • Server-side commits: every change to source history happens on the server, so nothing is pushed.
  • Partial checkouts: a job only gets the directories it asks for, and files are fetched as they are read.
  • Few branches: each branch is shared state that all servers must agree on, so Scion mostly sticks to main and release branches.

Anthropic says the new setup keeps up with its volume and uses 200x less CPU and memory than the fleet of Git mirrors it replaced. Arize notes that the ideas are not new: Google's Piper and Meta's Sapling have handled huge monorepos with similar tradeoffs for years. The difference is that those companies needed tens of thousands of engineers and years to reach that point, while Anthropic is much smaller and younger.

Antique, merge queues and the end of line-by-line review

Scion came with Antique, a code review system built at the same time. Arize says Antique can tell a comment-only change from a sensitive logic change, so humans review only the changes that justify their time. On the Scion codebase itself, engineers now review the intent of each change instead of reading every line.

Arize argues that this shift would have strained Git even without a giant tree, because merging becomes extremely frequent once human review is no longer the gate. At agent scale, hundreds of changes can merge while tests are still running. Anthropic could not make the usual global lock on the merge queue work, so it moved to testing after merge and before deploy. It accepts that concurrent changes will sometimes conflict, because Claude can fix those conflicts faster than a human can spot them.

Perplexity landed in a similar place, according to the post. It does not use pull requests at all, since no human reviews each change, and relies on automated checks when a commit is submitted.

Do other teams need to leave Git?

Arize points to companies that coordinate many agents without abandoning Git. Perplexity described the storage design behind its Projects feature, which it calls a Git forge for agents. Each agent works in a short-lived sandbox and submits commits to a service running stock Git on top of S3. If two agents collide, the one built on an outdated version is rejected, replays its change and tries again. Perplexity says fewer than 2% of commits collide. It keeps repositories small and squashes older history into a snapshot.

GitSwarm, a research system from researchers at Meta and several universities, has identical agents contribute to one shared, branching Git repository with no planner assigning tasks. The repository itself is the coordination mechanism, and it is ordinary Git.

Arize's conclusion is that Scion solves the problem of many agents working on a single giant monorepo, not the problem of having many agents. Anthropic itself says Scion is not a general replacement for Git. It is not releasing Scion as open source, partly for that reason and partly because adopting it would mean replacing code review, code search and CI too. The company says it is only necessary for a monorepo as big as its own, though it expects the problem to become more widely felt within the next 6 to 12 months. For most teams today, Git still works.

Related news

Coframe raises $35 million to build its self-improving ‘living internet’
BusinessNovember 9, 2026, 15:40

Coframe raises $35 million to build its self-improving ‘living internet’

Coframe has raised $35 million to develop an autonomous AI platform that continuously generates, tests and improves websites, landing pages and ads. The company says its…

Vercel adds streaming generative UI to Chat SDK agents
For developersNovember 6, 2026

Vercel adds streaming generative UI to Chat SDK agents

Vercel says Chat SDK agents can now stream generative UI with json-render, allowing agents to generate cards from typed JSON specifications. Chat SDK adapters then…

Business
BusinessNovember 5, 2026, 08:01

Sinch Q3 sales rise 5% organically as Americas drives growth and AI ties deepen

Swedish cloud communications company Sinch reported third quarter 2026 net sales of SEK 7,085m, an organic increase of 5 percent, with the Americas region cited as the…