Anthropic Built Its Own Source Control System (and the Controversy Behind It)

I read Anthropic’s article Redesigning source control at Anthropic and learned that Git had become a bottleneck once ai agents started writing and merging code much faster.

To solve this, Anthropic built Scion, a centralised source control system and Antique, a code review tool. Unlike Git, Scion keeps the code history on a central server and only downloads files when they’re needed. The server also handles things like file history and git blame so clients don’t have to.

For an easier overview with more industry context, such as how Google and Meta solved similar problems, read Gergely Orosz’s Anthropic builds its own source control system in The Pragmatic Engineer.

The controversy: giving credit

The part I found most interesting was the criticism of Scion. Rain, an engineer who worked on Meta’s Mononoke (the server behind Meta’s Sapling source control), pointed out that many of Scion’s ideas already existed in Mononoke especially its Bonsai data model and derived data.

Rain’s main issue wasn’t that Anthropic reused these ideas. It was that Anthropic’s original article didn’t credit the earlier work. After the criticism, Anthropic updated its article to credit Mononoke.

To understand the criticism in more depth, read An outsider’s perspective on Scion and Mononoke by typesanitizer, an engineer on Databricks’ source control team. They compare Scion with Mononoke and question some of Anthropic’s numbers. To see how other developers reacted, check out the Lobsters discussion of that post.

This stood out to me because reusing existing ideas is a normal part of software development. Still, I think it’s important to recognise the people who came up with them first especially when you present a new system as a major achievement.

What I learned

• Git blame: Shows who last changed each line of code and when which helps developers understand a codebase’s history
• AI creates new challenges: When code gets written faster tools like Git, CI and code review also need to keep up.
• There’s more than one way to do source control: Git works well for offline development while centralised systems can suit always online AI agents better.
• Giving credit matters: You don’t always have to invent something completely new but you should acknowledge the people whose work helped you.
• Look at both sides: Reading the criticism alongside Anthropic’s article gave me a much better understanding of what happened.

My main takeaway

This turned out to be more than a story about replacing Git. It also showed me how much credit matters in software engineering and building on other people’s ideas is completely normal but so is recognising their contributions.