Managing AI Agents Is Work. Hiring Plans Should Count It

Anthropic estimates that reading all the output from its interactive Claude sessions would consume more than half a working week even for nontechnical employees. That calculation uses text volume and an assumed reading speed; it does not measure time employees actually spent reading. [1]

I think this finding exposes a staffing question inside the promise of AI productivity: who has time to supervise everything the tools produce? My view is that employers should count supervision as part of the job and give workers authority to limit incoming work. Otherwise, a faster system could simply leave a person with a longer queue.

A growing task inside existing jobs

Anthropic reports that work on agent instructions, monitoring and related processes increased from 10% to 22% of the average employee’s interactive Claude sessions between fall 2025 and spring 2026. The estimate came from automated classification of 2,000 short session summaries. It measures a share of sessions, not working hours. [1]

That distinction matters. I would not convert those percentages into a claim about how much of everyone’s day goes to supervision. But the finding gives me reason to treat managing AI as a responsibility that needs explicit time, ownership and support.

For a task requiring approval, producing more proposals creates more potential decisions. I would want a staffing plan to explain how those decisions will be handled before counting every additional output as a productivity gain. Generating something and bringing it to a useful conclusion deserve separate attention.

Faster production can still mean more meetings

The company also reports that meeting time among technical individual contributors rose 16% over the preceding year. Its analysis acknowledges that company growth helps explain increasing meeting demands; the report does not isolate AI as the cause of that increase. [1]

I do not regard more meetings as automatic evidence of waste. A discussion can resolve competing priorities that no amount of additional drafting will settle. People may agree that a proposal is technically sound while disagreeing about whether their team should spend time on it.

My inference is that automation can leave the work of reaching agreement intact, even when preparation becomes much quicker. If lower preparation costs encourage more proposals, that coordination burden could grow.

This changes how I would judge an automation plan. Time saved preparing a document and time needed to settle what it commits a team to belong in the same calculation. A shorter first step does not guarantee a shorter working day.

Responsibility needs room to say no

To me, autonomy becomes meaningful when it includes control over which demands deserve action. Being responsible for an expanding stream of output is a different proposition from being authorized to reduce that stream.

Consider two hypothetical versions of an operations role. Both employees manage agents that propose schedule changes. One can stop unnecessary runs and defer low-priority requests. The other must respond to every proposal while meeting unchanged deadlines for existing duties.

I would regard those as materially different jobs, even if their titles and software were identical. The first includes discretion over workload. The second concentrates demands on the person expected to resolve them.

Anthropic’s findings do not establish that its employees lack such discretion. My concern is about the choices other employers might make when adopting similar methods. A credible staffing plan should identify who can postpone work, narrow its scope or decide that a proposed task adds too little value to pursue.

A career choice about the shape of the day

For job seekers, I think managing agents deserves consideration as a change in the nature of work. A familiar title could come to involve less uninterrupted production and more decisions about competing requests. That prospect may appeal to someone who enjoys prioritizing and coordinating. Someone who prefers sustained periods of direct creation may judge it differently.

Neither preference strikes me as a failure to adapt. The question is whether the actual responsibilities fit the work a person wants to develop.

That makes control over workload a useful consideration when deciding where to apply. I would pay attention to an employer’s explanation of how additional assignments get prioritized and who can postpone them. Claims about independence become more informative when they specify which decisions belong to the employee.

Someone searching after a layoff may have little bargaining power over these arrangements. I would use the available information to rank opportunities, without making perfect workload policies a prerequisite for applying. Missing detail should remain an unanswered question.

For skill planning, I would give attention to prioritization and capacity planning: how work enters a process, how much can be handled, and what happens when demand exceeds that capacity. Domain knowledge remains central to my reasoning. Deciding what can wait requires understanding the consequences of delay.

A scarce resource does not guarantee a vacancy

Anthropic’s report describes its own workforce using internal records and employee interviews. [1] I would be cautious about treating those arrangements as a forecast for every employer. The evidence also cannot tell me whether growing supervision demands will produce new positions or additional duties inside existing ones.

An employer facing limited human attention could hire, improve filtering or pursue fewer projects. My concern is that it could also leave the workload unresolved while continuing to celebrate rising output. The existence of necessary work offers no automatic assurance that someone will receive adequate time to do it.

My prognosis is cautiously positive about opportunities to improve these processes, but uncertain about how many distinct jobs will follow. Better tools could reduce supervision demands, while more ambitious projects could increase them. Workers who can organize that demand may become more useful without becoming immune to staffing cuts.

I would judge an AI hiring plan partly by whether it accounts for the time needed to reach decisions. The job I would favor gives its holder a say in what enters the queue, alongside responsibility for what comes out.