Picture a single shared codebase so busy that a chart of its activity looks like a cliff face.
Lines of commits stack on top of each other every few seconds, and nobody is pressing a button by hand anymore. The work is arriving faster than any human team could review it, and the plumbing underneath finally gave out. That breaking point is now public, dated, and measured in exact numbers.
What actually broke, and why old code tools could not keep up
The strain traces back to a simple mismatch. Version control systems like Git were built decades ago around the pace of human programmers, who might push a handful of changes a day. They assume a programmer opens a file, edits it, tests it, and submits one change at a time.
Once AI coding agents entered the picture, that assumption collapsed. Agents can draft, test, and submit changes continuously, across many projects at once, without needing sleep or meetings. So the volume of proposed changes stopped looking like a human workflow and started looking like a firehose of commits.
Each of those changes still needs to be stored, checked out, tested, and merged safely into one shared history. That is where the hidden cost piles up, not in writing the code, but in the bookkeeping around it. A tool designed for dozens of daily updates starts to choke when asked to handle tens of thousands.
Inside the surge, told in the company’s own numbers
The company at the center of this is Anthropic, the maker of the Claude chatbot and the Claude Code programming assistant. In a post on its engineering blog, the company laid out exactly how fast its internal pace of software changes grew this year.
In May, Anthropic averaged around 2,500 code changes merged per day. By September, that figure passed 30,000 per day, a tenfold jump in roughly four months, a pace Anthropic says became its biggest development bottleneck.
So what changed between May and September was not the size of the engineering team. It was how much of the writing, testing, and small scale fixing got handed to Claude itself.
The company says the assistant can now iterate on research loops and build new services at a speed it had never seen before.
The hardest numbers, and where they came from
This is not a projection or an outside estimate; it comes directly from Anthropic’s own engineering team, published on October 13 as part of a detailed technical account of the rebuild. The same post disclosed the daily strain on testing systems that most readers never see.
“The history of our main branch alone is almost 100 GB, and a compressed working copy is over 25 GB. We perform on the order of 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes, and research jobs. Fetched naively, that would be over 600 PB of git traffic per day.”
That single figure explains why a company would rather spend months building its own replacement than keep patching an older system.
The fix, called Scion, was not a small side project. We built Scion, an in-house version control system, and fully migrated our monorepo from git in five months. A team of seven people built both Scion and Antique (our code review tool), not to mention half a dozen related systems.
The payoff, by Anthropic’s own account, was steep. The Scion infrastructure footprint is now 200x smaller than the previous git mirroring infrastructure in terms of CPU and memory requirements, and CI and research jobs check out code two orders of magnitude faster, which frees up a sizable amount of capacity across the fleet; CI alone realized a 25% CPU capacity win.
Much like the way a single shared code library can carry outsized value across the entire internet, as detailed in this earlier report, Anthropic’s new system underlines how much invisible infrastructure now holds modern software together.
Where the record does not hold, and what still slows things down
None of this means every company needs a bespoke version control system tomorrow. Anthropic’s volume is extreme precisely because its own AI products generate so much of its own engineering workload, a loop that does not exist at most software shops yet. Anthropic itself says Scion is not meant as a general replacement for Git and only makes sense for truly enormous monorepos, not for a typical team or personal project.
Even Anthropic admits the new system has limits. The company did not claim Scion eliminates all friction, only that it moved the bottleneck further down the pipeline, from merging code to the builds and testing steps that still have to verify every change before it ships.
Critics outside the company have raised a fair question about oversight at this speed, since tens of thousands of daily changes make it hard to say whether human reviewers can meaningfully check each one before it reaches production.
That tension sits alongside a broader pattern already visible elsewhere in software, including how everyday tools like simple consumer devices now ship with their own fast release cycles, just at a far smaller scale than an AI lab’s internal monorepo.
What this means for the rest of the software world
Anthropic was careful to frame this surge as the beginning of a trend rather than a one time event, saying the latent demand for merging code is practically infinite and that every bottleneck it removes immediately leads to more code produced.
That framing matters because it suggests other companies racing to deploy coding agents at scale will eventually hit the same wall Anthropic just climbed over. Git has powered software development for two decades, and it is not going away for the vast majority of teams anytime soon.
But for organizations where AI agents are doing a growing share of the actual coding, the math Anthropic published offers a concrete early warning about what happens when the pace of change itself becomes the hardest problem to manage.
For now, the record stands as a snapshot of an industry still learning its own limits, measured not in a product demo but in the exhausted plumbing of a company trying to keep up with its own machines.
