Anthropic's Claude agents pushed code merges to 30,000 a day in September
Git was built for engineers who type at human speed. Anthropic's own coding agents pushed its source-control infrastructure past human-scale assumptions, forcing the company to build a replacement from scratch.
In May 2026, Anthropic averaged roughly 2,500 code changes merged per day into the company's internal monorepo. By September, that number had passed 30,000 a day, more than a twelvefold jump in four months. Anthropic disclosed the figures on its engineering blog, saying the surge reflected Claude's ability to write software at a scale that made source control, builds and CI the bottleneck.
The company said it had become "bottlenecked by developer infrastructure built for human, rather than agent, scale." Git, the version control system nearly every software company runs on, was designed around the assumption that a person sits at a keyboard, writes a batch of changes, and commits them every so often. Anthropic's monorepo stopped looking like that. According to the company, nearly all of its source-control clients are now agents and jobs that are online by definition, a pattern Git's original design from 2005 never anticipated.
So Anthropic built its own replacement. A team of seven engineers spent roughly five months designing and shipping a new version control system called Scion and a code review tool called Antique, while also migrating the monorepo off Git. Anthropic says Scion is not meant as a general-purpose Git alternative for other companies; it was built for enormous monorepos and swarms of agents, and its internal code review, code search and CI pieces are highly bespoke.
The gains Anthropic is reporting are not marginal. The company said Scion cut CPU and memory usage by roughly 200 times compared with the prior Git-based setup, and that code retrieval times for continuous integration and research tasks dropped from minutes to seconds. For a lab running constant automated builds and test cycles across a sprawling codebase, that difference compounds daily rather than showing up once.
Anthropic relaunched Projects inside Claude Code on September 17, adding a coordinating agent that splits engineering goals across multiple parallel cloud sessions called threads. Each thread works its own branch and shares memory with the rest of the project, though execution is cloud-only for now and the beta is limited to select Pro and Max... - how to use Claude Code multi agent projects - Claude Code parallel cloud agents for engineering tasks
There's a recursive detail worth sitting with: Claude helped build its own replacement plumbing. Anthropic said an early version of Scion had a prototype pipeline rewritten by Claude to run four times faster while using 20 times less memory. The tool that created the merge-volume problem also helped engineer the fix for it.
This isn't the only data point Anthropic has put out this year on how much of its own codebase Claude now writes. In a June 2026 essay, the company said that more than 80% of code merged into its production codebase as of May 2026 was authored by Claude, not human engineers. Put the two numbers together and the picture sharpens: Anthropic isn't just using Claude to speed up the humans who write code. It's running a development pipeline where agents are the primary actors and humans increasingly supervise and review.
Why Git actually buckled
It's worth being precise about what broke, because "Git can't handle it" undersells the mechanism. Git was built around discrete, infrequent commits from a relatively small number of human contributors, each working on a local copy of history that occasionally syncs with everyone else's. Anthropic's agents don't work that way. They run continuously, in parallel, often dozens or hundreds at a time, each generating small changes, tests, and merges on a clock that has nothing to do with a human workday. At that volume, operations that were fine as occasional background cost, repository indexing, history traversal, dependency resolution, turn into a constant tax that Git's architecture was never tuned to absorb cheaply.
Frankly, this is the clearest real-world evidence yet that frontier AI labs have outgrown the software infrastructure most of the industry still runs on. A twelvefold jump in merge volume in four months isn't a company getting marginally more efficient. It's a company whose primary workforce changed category, from people who type to systems that never stop running, and whose old tools simply couldn't keep pace.
Other AI labs scaling up agentic coding tools are watching the same math play out, even if their numbers haven't hit a press release yet. GitHub, GitLab and other version-control vendors built for human commit cadences will face a version of this same bottleneck as agentic coding spreads beyond the companies building the models. Anthropic's answer was to build its own system and keep it closed. Whether that becomes the standard response, or whether a general-purpose fix for agent-scale version control eventually ships for everyone else, is the open question Anthropic's blog post didn't answer.
ASML raises 2026 sales outlook for a third time as AI demand outruns supply • A jailbroken PlayStation 5 is now running a 27 billion parameter AI model • Seedance 3.0 Brings Feature-Length Control to AI Video Generation
Claude Code turned Anthropic from an admired safety lab into a company with a $47 billion run-rate, and it did it inside VS Code, the free editor Microsoft gives away and most developers already use. The breakout was a distribution story, and the platform underneath it belonged to Satya Nadella. - claude code vs code integration strategy - how microsoft enabled anthropic's growth
This article is posted in AI News, check it out for more related stories.
