
Redesigning source control at Anthropic
Simon Willison's link post on Anthropic's write-up about building Scion, a centralized source control system that replaces Git, and Antique, a new code review system. Anthropic's merged code changes rose from about 2,500 per day in May 2026 to over 30,000 per day by September, and the main-branch history is nearly 100 GB with about 25 million CI checkouts per day. A team of seven finished the migration in five months. Willison also notes Anthropic's use of adversarial review, which can push Claude toward over-engineering unless skilled engineers steer it.
What it claims
This is Simon Willison's link post pointing to Anthropic's own article; it is not the article itself. Anthropic's case for replacing Git with Scion rests on two scale figures the catalog summary omits:
- Working copy size. A compressed working copy is over 25 GB.
- Naive traffic. Fetched naively, about 25 million CI checkouts a day would produce over 600 PB of git traffic per day.
The 25 million figure counts CI only. It excludes developer environments, agent sandboxes and research jobs, so total checkout load is higher than the headline number. The driver is merge volume, which went from about 2,500 merges a day in May 2026 to over 30,000 a day in September.
How it works
The post gives almost no design detail. Willison reads Scion as "a reinvention of centralized source control along the lines of Subversion." That is his characterization, not a description quoted from Anthropic. Antique, the new code review system, is named but not described.
Evidence
Willison had Claude modify Anthropic's merge-rate chart to mark the 2026 public release date of each Anthropic model, and he links the code. He guesses that Fable/Mythos-class models are a key factor in the velocity jump. He flags that the markers show public release dates, not internal availability, so the chart cannot establish timing causality.
On adversarial review, Willison reports the same over-engineering tendency in his own projects. His setup plays Claude and GPT against each other to find flaws.
Limits
- Second-hand source. Everything about Scion comes through short quotes. Architecture, consistency model, Git compatibility, how the 600 PB is avoided, and Antique's design are not on this page. Read Anthropic's original for any of that.
- Commentary is opinion. The Subversion comparison and the model-velocity link are Willison's opinion.
- Narrow relevance. The figures matter mainly as a threshold. Unless your repo, working copy and checkout volume are in this range, nothing here argues for leaving Git.
Organisations named here
- Anthropic
- OpenAI
- Simon Willison
Technologies named here
- Claude
- Scion
- Antique
- Fable/Mythos class models
- GPT
- Git
- Subversion
Techniques named here
- Adversarial review
Related in coding-agents
