Anthropic published a technical report on October 16, How employment is changing at Anthropic, written by four researchers at the Anthropic Institute. It uses the company's own data on agent usage, code, hiring and employee interviews to show what happens to work at an organisation that uses AI agents more heavily than almost anyone else.

Agents now log more hours than people

The researchers compare two measures. Agent working time is the time an agent is actually running, including inference and tool use; waits longer than 30 minutes are excluded. Human working time is headcount multiplied by 40 hours a week. On that basis, autonomous and interactive agent hours together first exceeded human work hours in March 2026.

The headline figure is narrower. Counting only interactive sessions that employees start and direct themselves, with their subagents included, the agents worked about six times as many hours as their human handlers by August. The report also finds that about 80% of the time in a Claude session is spent waiting for human input.

Fewer juniors, different hiring

Since early 2025, the share of junior-level employees in technical departments has fallen from 21% to 14%. Among new technical hires, the junior share fell from 27% to 9%. Junior job postings dropped particularly sharply after December 2025. This happened while the company tripled in size over the past year, so technical hiring still grew in absolute numbers.

In an April 2026 survey, most managers said AI had changed the skills they hire for. About a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI. An automated review of interview scorecards found that mentions of candidates' creative use of AI tools roughly tripled for technical roles and quadrupled for other roles since 2024. For technical candidates, comments on autonomy and judgment grew the most, while comments on domain expertise declined.

The caveats

This is Anthropic studying itself, and the authors are open about the limits. Anthropic is an unusual company, and they warn that many of its practices cannot be copied across the economy. Tool adoption was not random: engineers planning big projects may have taken up agents for them, and fast business growth could raise output on its own. Several measures rely on automated classification, and the findings describe correlations, not proven effects of AI. The authors also say AI could displace workers, and they call for public policy on retraining and income support.

What this means if you're learning AI tools

The skills Anthropic now looks for are the ones agents don't supply: setting a clear goal, working independently and judging the result. To build them, practise handing a whole task to an agent such as Claude Code and reviewing what it returns, rather than only asking a chatbot questions. For work that isn't code, an assistant such as Claude is a lower-stakes place to learn how to brief an AI, check its answer and correct it. The report's own data makes the point: agents spend most of each session waiting for a person, so knowing what to ask for and how to check it remains the human part of the job.