Claude now writes code so fast that Anthropic had to rip out git, the tool basically every software company on Earth uses to track changes to its code. Back in May, Anthropic was merging about 2,500 code changes a day. By September it was over 30,000. That's one every three seconds, all day and all night, with Claude driving most of the increase. Git was never built for that. It was designed in 2005 so every developer's copy of a project carries the full history and works offline. That's great if you're a person coding on a plane. Anthropic's main code history is now almost 100 GB, their test systems pull a copy of the code 25 million times a day, and nearly all of those requests come from AI agents and automated jobs that are always connected anyway. Done the normal git way, that adds up to 600 petabytes of traffic a day. They tried every standard workaround first. At one point they were running several hundred internal git mirrors just to keep up, and it kept falling behind. So they built their own system, called Scion. The server does nearly all the heavy lifting (commits, merges, diffs, figuring out who changed which line) and most jobs just ask it questions over an API without downloading the code at all. The new setup uses 200x less compute and memory than the mirror fleet it replaced, and checkouts that used to take minutes now finish in seconds. Seven engineers built it. Those same seven also built a new code review tool and code search, then moved the entire company over in five months, while the amount of code flowing through was growing 10x under their feet. Google spent years building its version of this, called Piper, and Meta put the better part of a decade into scaling Mercurial into what's now Sapling. Anthropic openly says it borrowed lessons from both. Still, seven people. Every engineer on that team now says Claude writes code better than they do. Early on, they took the ingestion pipeline they'd mostly written by hand, gave it to an internal preview of Claude Mythos with a request to rewrite it and no other instructions, and got back a version that used 20x less memory and ran 4x faster. So what were the humans actually doing? Making the big design calls, and reviewing each change at the level of what it was trying to accomplish. And according to the post, a lot of the job was telling Claude "no." The model would sometimes bend core system rules to make a bad pattern work, or pile on defensive code for problems that were never going to happen. They even gave up on the traditional merge queue. When thousands of changes can land while a single test run is still going, merging one at a time with the whole repo locked just stops working. So they test after merging and let Claude untangle the conflicts, which it usually does before a human would even notice them. Nearly every tool software teams use was built on the assumption that a person types every line. Anthropic hit the wall first, and they're saying they expect plenty of other companies using AI coding tools to hit it within the next 6 to 12 months.