Anthropic’s agents outrun its people

AI agents at the company work six hours for every hour people put in.¹ They produce more than anyone can read, so human attention sets the pace. But checking that output needs judgment, learned in the junior jobs AI takes over first. So it’s back to square one. The bottleneck is people.

The figures come from a report Anthropic based on its own records of hiring, agent use, output and meetings.¹ Nobody outside the company can check them, and the authors work for the firm whose tools they measure. Anthropic also tripled in size in a year, so growth, not just AI, lifted the numbers. The authors try to separate the two by following the same employees before and after they adopted agents. On its own, the report is one company’s view of itself.

However, outside studies find the same human bottleneck and the same drop in junior hiring.

The agents wait for people

Output per person did rise. The average engineer at Anthropic now submits five times as many code changes as six months ago.¹ That output still lands on people’s desks. When an employee works with an agent in a live session,

the agent spends about 80% of the session waiting for the employee to respond.¹

For engineering and research staff, reading one week of agent output would take over ten working weeks.

Nobody reads it all. Staff build tools and write instructions so that agents check other agents. That work has grown from a tenth to over a fifth of their AI sessions.¹

Developers elsewhere hit the same wall. In one study, those using autonomous coding agents submitted about 240% more code changes but released only about 30% more software.² Its authors put the gap down to human bottlenecks.

AI was supposed to free up people’s time.

At Anthropic, the median employee now spends 20% more time in meetings than a year ago, and technical staff without management roles 16% more.¹

With routine work handed to agents, what remains is strategy, and strategy needs people to argue it out. Chasing bugs, too, increasingly turns up problems in how the company is organised, and people settle those together.

Judgment has to come from somewhere

Checking agent output takes judgment, and Anthropic now hires for it. In interview notes for technical roles, autonomy and judgment come up more often than before, and domain expertise less. For other roles, execution comes up less.¹

Judgment, though, comes from years of writing code, drafting documents and checking details. That work used to belong to juniors, and agents now do it. At Anthropic, juniors fell from 27% to 9% of new technical hires.¹ Across the US, young workers in the jobs most exposed to AI are losing ground, mainly because fewer of them are being hired.³ Experienced workers in the same jobs are holding steady. A job that never opens never shows up in layoff figures, so the shift is easy to miss. A company can hire experienced people only while someone else is still training them. If every company makes the same choice, nobody does the training.

The sociologist Manuel Castells drew a line between two kinds of work decades ago. He set people who choose their own course towards a goal against people who carry out tasks set for them. Tasks set by someone else can be written down, and written instructions are what AI models run on. A graduate arrives mostly with written knowledge, which the models already hold.

An experienced worker brings know-how nobody has written down, and Anthropic is writing that down too, in thousands of standing instructions for its agents.¹

Some of its employees are already drifting away from the hands-on work that built their own skills. The authors admit that nothing the company does directly addresses the slowdown in junior hiring.¹

Anthropic is an extreme case

One employee describes constant pressure to use AI for everything.¹

A company that automated only the repetitive part of a process would give people less to check. The report leaves that path untested. The bottleneck stays with people either way. How much output reaches people, and in what form, is a design decision. Where the next generation of experienced people will develop their judgment is still an open question.