Anthropic has stopped hiring junior people at anything like its old rate. They were 27% of its technical hires in early 2025, and the share is now 9%. Over the same time, the typical employee's meeting time went up by 20%. Anthropic is an extreme case, an AI company that has rebuilt its work around Claude and tripled in size in a year, but the two changes have one cause, and it will reach other workplaces.
The scarce thing is now human attention
At Anthropic, Claude agents now work six times as many hours as the employees who direct them. Most of the time an agent spends in a session is spent waiting for a person, about 80% of it. And what the agents write piles up. Even for staff outside engineering, reading everything their Claude sessions produce would take more than half the work week. At normal reading speed that is about 300,000 words, roughly three novels, every week.
Nobody reads three novels a week, so people stop reading and start building ways to avoid it: dashboards, standing instructions, agents that sort other agents' requests. That kind of work grew from 10% to 22% of employees' Claude sessions in about half a year.
Here is how it could look in an ordinary business. An operations manager at a regional distributor gives an agent the weekly supplier invoices. Before, a junior analyst matched them to purchase orders over two days and brought the manager the dozen that looked wrong. Now the agent matches all of them in minutes and writes a note on every one. The manager's time has not gone away. It has moved from checking one analyst's work to deciding which of 300 notes deserve a look. The most useful thing the manager can do is tell the agent to bring only the exceptions.
Meetings are where the remaining decisions get made
Once the doing is cheap, what is left is deciding what to do and getting people to agree. That needs people talking to each other, and the extra meetings follow from it. Engineers are in meetings 16% more than a year ago. One account executive let Claude handle email drafts and record keeping, and used the freed time to talk to 30% more customers.
There is a practical twist for small firms. An agent can only work from what it can see. If a decision is made in a meeting and nobody writes it down, the agent keeps working from the old plan. At Anthropic, one employee's name on a months-old document kept drawing agent requests for t-shirts and mugs long after that person's job had changed. So the notes from a meeting have become instructions to your software, and a meeting with no notes now leaves your AI behind.
The first job is the one that disappears
Hiring managers at Anthropic now write about autonomy and judgment in their interview notes more often, and about specific technical expertise less often. That fits the junior numbers. The entry job was mostly doing, and the doing is the cheapest part to hand over.
The bank tellers show why that matters. Their job survived because cash and customers were in the same seat, so the human part came attached to the work. A junior engineer's code had no human part attached. Employers now want judgment from the first day, even though judgment is what used to be learned during the doing. We looked at that learning problem in September, in how AI weakens the way young workers learn on the job.
A manager can respond directly. Keep one or two junior seats, but rebuild them: the junior reviews what the agents produce, sits in the meetings where decisions are made, and explains to a senior why they approved or rejected each piece. That is the apprenticeship that used to come free with the typing. A company that stops hiring juniors saves a salary now, and has to buy judgment on the open market later.