When agents scale, human-sized systems break first (Anthropic)
Anthropic's merged code changes went from about 2,500 a day in May to more than 30,000 a day in September. Git, the tool almost every software team on earth relies on, could not keep up. So Anthropic built its own version control system, and a seven-person team migrated the entire monorepo off git in five months.
The breaking point was ordinary plumbing. Queues, locks and merge conflicts that are minor at human volume become the system itself at agent volume.
The lesson is not about software engineering. It is about what breaks first when agents arrive at scale.
It is rarely the model. It is the surrounding system, designed around human throughput: approval queues, review steps, ticketing, month-end close, document management, procurement workflows. All of them quietly assume a person produces a few units of work an hour. Multiply volume by twelve and those assumptions fail, often silently.
There is a counter-case worth naming. Anthropic is an extreme case: an AI lab generating code at a volume few companies will ever see. Most organisations will not need to rebuild their tools. But the direction travels. The first system to strain tells you where your controls were sized for people.
Most enterprise AI programmes are still asking which tasks agents can do. The better question is which of our systems become the bottleneck when they do. The answer is usually governance infrastructure: the controls, audit trails and sign-offs that were never load-tested for machine volume.
The same shape shows up outside engineering. An agent that resolves ten times the tickets still feeds one complaints queue. An agent that drafts every reconciliation still feeds one review step.
You can redesign those systems deliberately, or watch your teams route around them. The second is how controls die quietly.
If your work volume rose tenfold next quarter, which process would fail first? That is where your AI strategy starts.
#OperatingModel #Transformation #AusBiz
Source: Anthropic engineering, 13 October 2026: https://lnkd.in/gMamyPvr