The Information Machine
Following·Day 4·in the edition since 14 Oct 2026·4 sources·updated 16 Oct 2026

Anthropic's Scion replacement for Git

The gist

Anthropic Built Scion VCS as AI Code Authorship Rose From 3% to 26%

Anthropic's own infrastructure had to be rebuilt to handle AI-agent-driven code volume, and the jump in AI-authored code share from 3% to 26% within months illustrates how quickly agent-scale workflows can outpace tooling designed for humans. The concrete operational metrics give a rare window into what AI-accelerated software development looks like at a frontier lab.

The full picture

Anthropic replaced Git with an in-house version control system called Scion after daily code merges grew from roughly 2,500 in May 2026 to over 30,000 by September 2026, a surge Anthropic attributed to AI coding agents rather than human developers. During the same period, AI-authored code at Anthropic grew from roughly 3% to 26% of R&D output. A seven-person team built Scion and completed the migration in five months. Scion resembles centralized systems like Subversion rather than distributed Git. Anthropic also built a companion code review system called Antique. The scale that drove the switch: the main branch history alone reached roughly 100 GB, and with approximately 25 million CI checkouts per day, naive Git fetches would have generated over 600 petabytes of traffic daily. Anthropic said the prior infrastructure "became bottlenecked by developer infrastructure built for human (rather than agent) scale." An early internal Claude model rewrote an ingestion pipeline without further instructions, producing a version 4x faster with 20x less memory. Anthropic also uses an adversarial review technique where one AI model critiques another's finished work rather than approving it, though the approach can turn counterproductive when the reviewer fixates on a narrow concern and causes unnecessary complexity.

How it developed
15 October 2026

AI-authored code at Anthropic grew from 3% to 26% of R&D output during the Scion development period, with a Claude model rewriting an ingestion pipeline without instructions to run 4x faster with 20x less memory.

October 15 reporting also described Anthropic's adversarial review technique, where one model critiques another's finished work rather than approving it, noting the technique can turn counterproductive when the reviewer fixates narrowly, which Anthropic called 'myopic'. Rob Bensinger argued it is 'time to panic about AI,' suggesting loss of control could be years or months away.

First in the edition · 14 October 2026
13 October 2026

Anthropic published an engineering blog post on October 13, 2026 describing its replacement of Git with an in-house version control system called Scion, after daily code merges in its monorepo grew from roughly 2,500 in May 2026 to over 30,000 by September 2026, a surge Anthropic attributed to AI coding agents.

The existing infrastructure, Anthropic said, "became bottlenecked by developer infrastructure built for human (rather than agent) scale". Seven engineers built Scion, which resembles centralized systems like Subversion, along with a companion code review tool called Antique, completing the migration in five months. At roughly 25 million CI checkouts per day, Anthropic said naive Git fetches would have generated over 600 petabytes of traffic daily.

Sources
The daily email

Want this in your inbox?

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free