The bottleneck has moved from building to reviewing

At the companies furthest ahead, AI agents already put in more hours than the staff who run them. The scarce resource is now human attention, and that changes how software should be bought.

For most of the history of software, the expensive part was writing it. That is changing faster than most budgets have noticed. In October, Anthropic published a look inside its own workforce: by August, interactive AI agents were putting in six hours of work for every hour its employees did, and output per person had risen across engineering, sales and contract work.

The same report found that the limit had moved. Reading everything their agents produced would take even non-technical staff half of their working week, so people were spending more of their time on the work around the work: building dashboards, watching pipelines, deciding what to trust. Hiring shifted too, with fewer junior technical roles and more weight on autonomy and judgement than on hands-on implementation skill.

The frontier and the market are moving at different speeds

A companion survey shows how far ahead that experience is. In August 2026, Anthropic polled 1,334 of its own staff and, through Morning Consult, 10,980 US adults. On average, staff expected AI to do 75% of their current work hours by 2030; US workers put the figure at 28%. Among people who use AI at work, 59% of Anthropic staff said some task now takes a quarter of the time or less, against 26% of US adults.

Staff at an AI lab are not a neutral sample, and these are expectations rather than measured outcomes. But the gap is the useful part. Most organisations still budget as if building is the slow step, while the teams at the edge are already constrained by something else.

What it looks like outside a lab

The Next.js team gave a concrete example in September. Their issue tracker held more than 2,000 open reports, many of them already fixed, duplicated or out of date. They built a research agent that investigated each report in an isolated sandbox, tried to reproduce the bug on current releases and returned its evidence, but could not close anything itself. Maintainers reviewed that evidence and, in about three weeks, 1,462 issues were closed and the backlog fell below 1,000.

That is the pattern worth copying. The agent does the reading, reproducing and cross-checking. A person makes the call, and makes it quickly, because the evidence is already gathered.

What this means if you are commissioning software

• Building is getting cheaper; deciding is not. The scarce work is knowing what to build and confirming that it does the job.
• Review capacity is the real constraint. Budget for tests, evaluations and dashboards the way you used to budget for developer hours.
• Put evidence in front of decision-makers early. A working prototype settles in an afternoon what a specification argues about for weeks.
• Keep a person on the irreversible steps. Let agents research and draft; let people approve, merge and send.

None of this makes software free. It moves the cost to the places where judgement lives, which is where a good studio should be spending your money anyway.