Anthropic Completely Rebuilt Git (Really)

Anthropic invented their own version control system, reinvented Git, and built it for themselves. Command A And it's basically because they are using agents so much, they're pushing so much code up into Git that they just needed a whole new system. So, they called it Scion, but if you write code with a especially with a team, then you've got your main bit of code that goes to production, and every time you want to build a feature, you duplicate the code, make all your changes, and then you make a request for that code to be put into the main production branch. And Git kind of tracks all of the changes, the text changes, and two of you change the same line of text, then that's what they call a conflict. So, it's good for humans, and as they say in this article here, it's still good for most people. Git's, right? This is something very bespoke to them, and it's not something they're going to be releasing. Going through the article, in May 2026, Anthropic averaged around 2,500 code changes merged per day. As of September, we've merged over 30,000 per day. That's crazy. Um, what I've noticed, though, is that I'm updating my apps all of the time now. I'm finding there's just always an update for an app, and it's quite tiring. Like, I feel like they need to wait until they've got some decent chunk, cuz it's just it's [ __ ] update fatigue I've got going on here. To in order to support this growth, we built Scion. We an in-house version control system, fully integrate fully migrated our mono repo from Git in 5 months. They started building it when they launched Fable, Mythos, if you remember back then. That's when they started with it. Anthropic [snorts] uses Claude to write software far more than most almost any other company. We're so often the first to see where tools built for human scale start to strain. Git Git is a decentralized system, as I explained. Every clone is a complete self-contained repository that can compute diffs, read history, blame, and check out any revision while still fully offline. So, the the big idea behind this is that the properties that make Git excellent for distributed offline work were actively undermining our ability to scale as the repo grew. So, we asked ourselves, if we design the source control with the fact that agents are always online in mind from the start, what would that look like? So, that's the component to this, that the everything's online, you're always connected. And there are a few little workflows where I've begun to think about like that. Like, I'm always online. This computer is always online, so what can I move to that to just do stuff for me, cuz I just know that it's always going to be online. Every client is capable of being a full-fledged server. In Scion, the server answers questions about the file history and blame. So, it's a it seems like it's a wrapper around it, that you don't ask the file directly, you ask the service in front of the code, and it tells you. Do you make branches? Do you make work trees? Do you use GitHub to Do you use merge requests and things like that? Like, what's your workflow? Very minimally. I I do use it and even with some teammates, but I'm not a pro, right? Quite new to it, so we are using it. I'm I'm doing the merges, uh like Mhm. but I don't actually like read the code. I I look at what this is or like talk to my people about it, and then hit the merge. Um but that's kind of like the extent. I do use work I did use work trees, but nowadays much less, because it's very rare for me work on the same file, you know? Like, sometime like very often I'm making a component while doing the login side of things on the another session, so the agents, they don't, um you know, work on the same thing at the same time. Yeah, so it's very interesting like lately tell Claude to I said like push main then it comes back and says you know, there is another agent they are not finished with their work. Let's wait on them or like when that one is finished tell me and I will push everything including their commits main. So it seems like they are very aware of the other agents working on the other files and that works out really well or even if the other agents are done and you're like push domain it tells you you know what actually there are other commits from another session. Trust me if you're pushing straight to main you're definitely pushing bugs up into main even with fiber or even with opus right? They're still going to there's always an issue. So never is there a perfect bit of code that goes in so I do it for the code reviews. And the only other thing that I marked in this article which was quite interesting and we've kind of already mentioned it is like how merge requests are solved. Basically there is an agent and wrapping all of this like we we said there's like an API layer that's answering questions. I think it's literally an agent there is just this master [ __ ] thing taking everyone's code changes working them out and this and that managing conflicts. So if two people edit the same file it looks at those two edits and figures out what needs to be done. So they put a lot of faith in their agents. I swear to God they're probably using a beast mode version they're probably using opus six or something like that. Huge undertaking like I say it won't be something that we can use I mean probably not we it doesn't make sense for us to use it in this way kill really but really impressive thing that's taken them five months to do and I like that they are committed to being very transparent about this cuz like I say this would this is never going to get released but they are committed to making sure that they are staying abreast of the dev community. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.