Anthropic admits its agents wait on humans 80% of the time: we are now the bottleneck

By Fast Company · October 19, 2026.

Anthropic has published a report on AI's impact on its own workforce. According to the company, its Claude agents spend 80% of session time waiting for human prompts rather than actively working. Anthropic says this is the cause of the productivity bottlenecks it has found.

The company says internal use of Claude translates into "more code per person, more deals closed per salesperson and more contracts processed." It also argues that its internal trends "could be a leading indicator" of how the organization of work, hiring and labor demand will change across the economy. According to Fast Company, that points to fewer entry-level jobs and a workforce focused on supervising agents.

The human data is the most striking. In an April 2026 survey, 75% of employees said the company's work processes make them productive. In an open-ended question, 20% said their role has shifted "from doing the work to managing Claudes." One employee compared the job to babysitting: every agent demands attention with the same urgency and can't tell a trivial problem from a serious one ("I need you to open my juice" versus "I've cut my finger").

The article notes that the report comes weeks after an employee resigned and accused the company of racing toward a "self-improving superintelligence." The accusation is that person's, and Anthropic has not confirmed it in the material available to us.

**Our take.** The 80% figure is less paradoxical than it seems. If an agent does in minutes what would take a person hours, it's normal for it to spend most of its time idle, waiting for decisions. That doesn't show that the human is in the way. It shows that value is shifting toward deciding what matters, in what order and by what criteria. And that's where the employee's problem lies: agents still don't know how to prioritize.

The report should be read with caution. It comes from a company that sells these tools measuring its own success, with a highly skilled workforce as its sample. In general, extrapolating from an AI lab to the rest of the economy calls for prudence: the average employee at an insurer or a law firm doesn't have the same autonomy or the same tools.

Even so, the pattern fits what we have been arguing about jobs: administrative and repetitive work is the most exposed, and what survives is judgment. The uncomfortable part, in the short term, is the entry-level rung. If agents absorb junior tasks, how will the people who must later supervise them be trained? That is the question companies and education systems should be answering now.

In the long term, our take remains constructive. Managing poorly calibrated agents is a transitional phase: as they get better at prioritizing and at asking for help only when it matters, the 80% waiting time will fall and supervision will become lighter. What feels like "babysitting" today is the cost of learning to delegate to an immature technology, and it's a reasonable cost if the destination is freeing up human time for what really matters to us.