Marcus: Anthropic is merging thirty thousand code changes a day.
Tessa: Thirty thousand? A day? That's not a company, that's a small country's entire GDP of code. How?
Marcus: It's not people. It's their AI, Claude. It's just... writing software. Endlessly. And that firehose of code completely broke Git.
Tessa: Okay, so this is just the classic big company monorepo problem, right? Like what Google and Meta ran into. They have so much code in one place that the tools get slow. They just need bigger servers, more mirrors.
Marcus: That's what I thought too! But it's actually a different problem. This new engineering post from Anthropic lays it out. They had hundreds of mirrors. The issue isn't just the size of the code, it's the user. Git was built for humans, working on their laptops, often offline.
Tessa: And Claude is not a human on a laptop.
Marcus: Claude is an always-online agent. It doesn't need an offline copy. It doesn't need the entire 100-gigabyte history of the repository just to change one line of code. The whole philosophy of Git—decentralized, every clone is a full copy—is the exact opposite of what they need.
Tessa: I'm not sure I follow. A checkout is a checkout. What does 'always online' actually change about the process?
Marcus: It changes everything. They built a new system, called Scion, based on this idea. The client—the thing running the AI agent—is basically dumb. It holds almost no data. If the AI needs to know who last edited a file, it doesn't look at a local copy. It just asks the main server.
Tessa: So all the work happens on the server? The diffs, the merges, the history searches?
Marcus: All of it. They call it client-server asymmetry. And because of that, a lot of the time, the AI doesn't even need to check out the code to disk. If a job just needs to find the difference between two versions, it can just ask the Scion API, get the answer, and move on. No files copied. That's how they handle 25 million of what they call 'checkouts' a day.
Tessa: Wait, you said commits happen on the server, too. So there's no 'git push'?
Marcus: No push! A commit is just an entry in a log on the server. It's basically just a note that says 'these files changed in this way,' with a pointer to the new content. It's not creating this whole complex tree structure like in Git. It's way, way cheaper.
Tessa: So how does it do something complicated, like figure out who to blame for a bug in a specific line?
Marcus: So this is the part where it gets a little fuzzy for me. They say they build these 'derived data' structures from that simple log of changes. So they have separate, optimized indexes just for blame, or just for path history. Honestly, how they compute that recursively... I don't fully get it. But the point is, it's not part of the core commit, so making a commit stays incredibly fast.
Tessa: Okay, this is technically fascinating. It's also completely and utterly irrelevant to me and every single developer I know. This is a solution for a problem that only Anthropic has. It's an beautiful, bespoke engine for a car that doesn't exist.
Marcus: For now! But that's the whole point of their post. They're saying, 'we're just the first ones to feel this pain.' Any company that starts using AI agents to write code at scale is going to hit this exact same wall. This isn't a one-off, it's a preview of the next five years of infrastructure.
Tessa: Is it, though? This feels like building a Formula 1 car. The engineering is incredible, but you're not going to find that tech in a Honda Civic. They made a centralized system. The whole power of Git is that it's decentralized, which is what you need for open source, for human teams spread across the world. They traded all that away.
Marcus: That's a fair point. You're not going to download and install Scion for your side project. But the principles might trickle down. The idea of an API-first source control, where the server is smarter and the client is dumber... that feels like it has legs beyond just Anthropic.
Tessa: Maybe. Or maybe we'll just figure out how to make AI agents work better with the tools we already have. It feels like we're watching them build a cathedral for a congregation of one.
Marcus: A congregation that is writing 30,000 psalms a day. I don't know if this is the future, or just a beautiful, custom-built spaceship for a planet with a single, very busy inhabitant. But the problem they're solving is definitely coming. I'm Marcus.
Tessa: I'm Tessa.