An engineer at Anthropic recently described a normal working day as facing “an army of Claudes, with every one asking for my attention”. Anyone who has started handing real work to agents will recognise the feeling: the drafts arrive faster than we can read them, and the queue of things awaiting our say-so grows quicker than the work itself.

The line comes from a report the company published on 16 October about how employment is changing inside Anthropic. By August, the hours its interactive AI agents put in outnumbered its employees’ hours six to one. Yet the agents weren’t the ones setting the pace. Around 80% of the time in a Claude session was spent idle, waiting for a human to read something, answer a question or approve the next step.

Once agents do the work, the scarce resource in a business is the attention of whoever has to decide about it — and in most small businesses that person is the owner. So the more useful question for those of us running one is how much of our attention each agent needs, and how to make that number smaller. It’s one study of an unusually agent-heavy workplace, and the authors say plainly that much of it can’t be replicated across the economy; a five-person accountancy practice won’t be running hundreds of agent-hours a week next month. Small businesses that hand more of their work to agents are likely to meet the same constraint, though, just on a smaller scale.

Gerber’s owner, back at machine speed

This is an old problem in new clothes, and Michael Gerber described it better than anyone in The E-Myth Revisited three decades ago. The owner who does the technical work, holds every decision and signs off on everything becomes the ceiling on the business. Nothing moves until it has crossed their desk, so the business can only grow as fast as one person can think.

Vertical Leap ran on Gerber’s principles for the seventeen years I led it, and a good deal of that work came down to taking my own name out of boxes on the org chart, one role at a time, until the business could run without me. Last December I argued that Gerber was right but that AI changes what his advice means. The Anthropic report shows one reason why in unusually concrete terms.

Agents bring the trap straight back, only faster. A human employee who’s stuck will usually try something, ask a colleague or get on with another task while they wait for us, whereas an agent set up to wait for approval simply stops. Give an owner a dozen of them and we’ve rebuilt the bottleneck Gerber spent a book trying to dismantle, except that the workers now finish in minutes and all of them are queuing at the same door. The report gives a sense of how much reading that means: for Anthropic’s non-technical teams, simply reading everything their agents produced in interactive sessions would take more than half the working week, which leaves precious little of it for the job they were hired to do.

Where the scarce thing has moved

For most of business history, the expensive part of getting something done was doing it. We hired people because the work needed hands, and we managed them because the hands needed directing. Agents are making the doing cheap, which moves the cost upstream to deciding: what should be done, whether the result is good enough and what happens next. The company’s figures show the shift. The share of interactive sessions spent on what the authors call meta-work — building agent skills, reviewing output, keeping an eye on automated pipelines — more than doubled in about six months, from 10% to 22%. Its hiring managers are writing more about candidates’ autonomy and judgement and less about their domain expertise. Attention, not compute, is now the constraint, and the skill worth building is designing work that needs less of it.

What this says about the jobs worry

Pew Research found in June that 71% of American adults expect AI to mean fewer jobs over the next twenty years, and the worry is understandable. Yet in one of the most agent-heavy workplaces with published data, the machines spend most of their time waiting on people. That’s why, I think, the “jobpocalypse” people keep predicting hasn’t arrived in the shape they expected: the work is moving towards judgement and approval, both of which still need a person. It doesn’t prove anyone’s job is safe, and I’m less sure what happens to the junior roles where people used to learn judgement in the first place, given that the junior share of Anthropic’s technical hires fell from 27% to 9%. For an owner today, though, the evidence points to a practical job rather than a threat, which is making each person’s judgement go further.

Four ways to need less of yourself

Gerber’s remedies still hold up. They just need sharpening for workers that never tire of asking.

  1. Settle decision rights in advance: write down what each agent may do without asking you, such as sending the follow-up email, refunding anything under £50, reordering stock below a set level, or publishing to the drafts folder but never live. Every permission granted ahead of time is an interruption you’ll never receive.
  2. Write briefs good enough that the work doesn’t come back: a fair share of the questions agents ask are ones we could have answered at the start. When the same clarification comes up twice, it belongs in the brief or the standing instructions rather than in your inbox.
  3. Batch your approvals: being pinged several times an hour is the worst of both worlds, because the agent still waits and you never get an hour of uninterrupted thinking. Set two or three windows a day to clear the queue, and let the questions line up for those slots.
  4. Measure how long work waits for you: every vendor will tell you how fast its agents work. The number that matters in your business is how long work sits waiting for a human, and who that human is. If it’s mostly you, you’ve found your bottleneck.

A week of noticing

None of this needs a new tool, and the simplest way to start is to watch yourself for a week. Keep a note every time an agent or AI tool is waiting on you, and for what: an approval, a missing fact, a judgement call, a “does this look right?”. By Friday you’ll have a list, and the recurring entries on it tend to fall into two piles. Some are decisions you could make once, in advance, and hand over for good. The rest are briefs you could write better, so that the question never gets asked in the first place. Every decision we design once is attention we get back, and spending it on the calls that really do need us is about the best use of an owner’s day I can think of.