Anthropic has disclosed how AI agents are changing its own workforce, and the numbers are striking: the share of entry-level workers among new technical hires fell from 27 to 9 percent. At the same time, the agents that employees direct now put in many times more working hours than the humans themselves. The report is a rare inside look at a company that adopts AI earlier and more intensively than almost any other. It shows which skills are in demand there and raises a question that reaches far beyond Anthropic: where will tomorrow’s experts come from if the entry point disappears?
Key takeaways
- The junior share in Anthropic’s technical departments fell from 21 to 14 percent since early 2025, and among technical new hires from 27 to 9 percent.
- In August, employees’ agents worked six times as many hours as the employees themselves. Agent hours first exceeded human hours in March.
- Engineers produce five times as many code commits as six months earlier. The report mentions no layoffs.
- In hiring, tool fluency, autonomy and judgment count for more than pure subject knowledge.
- Anthropic itself stresses that a single, extremely AI-intensive company does not reflect the wider economy.
What the report measures
The report “How employment is changing at Anthropic” was released on October 16 as a technical report from the Anthropic Institute. Its authors are the economists Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. They analyze internal data on employment, AI usage, work output, calendars, Slack communication and hiring decisions, supplemented by surveys and 17 interviews with employees.
The headline figure on agent hours needs some explanation. The count covers agent runtime, including compute time and tool use, but excludes idle stretches longer than 30 minutes. Human hours are set at a flat 40 per week per person. The six-to-one ratio refers to sessions that employees steer interactively. In other words, it describes how much machine work one person sets in motion, not how many people are being replaced.
Fewer entry-level hires, different work
The technical departments kept growing in absolute numbers, but more slowly than sales, finance or recruiting. Their share of the workforce fell from 61 to 56 percent within a year. Within engineering, the weight is shifting toward experienced staff: postings for junior roles have dropped particularly sharply since December 2025. In an April survey, about a third of machine learning managers and a fifth of software engineering managers said they were hiring fewer people because of AI.
The work itself is changing too. The share of so-called meta-work, meaning setting up agents, monitoring them and maintaining their tools, rose from 10 to 22 percent of interactive sessions. About 80 percent of the time in such a session, the agent is waiting for human input. Median meeting time rose by 20 percent. In 2026, employees with at least one year at the company switched departments about 50 percent more often each month than before. One employee in the report describes how the strategic part of the job, once a fifth of it, now makes up almost all of it.
The finding on quality stands out: according to the report, agent-written code was not deleted or reverted more often than code humans wrote themselves. The researchers concede, however, that such proxies measure quality only roughly and that AI errors are often harder to spot.
What early-career workers can take from it
For students and career changers, the report contains concrete pointers. In Anthropic’s hiring interviews, fluency with AI tools now carries much more weight, roughly three times as much for technical roles and about four times as much for non-technical ones. At the same time, autonomy and judgment count for more, while pure subject knowledge and execution count for less.
- Learn to direct agents, not just use them: break down tasks, check results, build workflows.
- Show that you can verify: someone who spots when an agent is wrong is harder to replace than someone who simply writes code fast.
- Think broadly: sales, legal and accounting at Anthropic also work with coding agents. Programming is becoming a cross-cutting skill.
The case of Kenya’s ghostwriters showed how quickly AI can upend an entire line of work.
The commons question
The bigger picture comes from researcher Nolan Lovett of the NATO Special Operations University. In his paper “The Tragedy of the Cognitive Commons,” published in July in the journal Human Resource Development Review, he applies the classic tragedy of the commons to professional expertise. A company that replaces entry-level jobs with AI saves money right away. The costs fall on the whole industry, because expertise has traditionally grown in exactly those jobs. Lovett also points to a trap: whoever is supposed to oversee AI needs the very expertise that could be lost when entry-level jobs disappear. According to The Decoder, he expects the full consequences between 2030 and 2045, especially in software development, financial analysis and legal research.
Whether Anthropic is a harbinger remains open. Co-author Peter McCrory sees the company about 18 months ahead of the rest of the economy in AI adoption. Broader research is mixed: a Stanford study led by Erik Brynjolfsson found markedly fewer hires of young workers in AI-exposed occupations, while a Danish study found no difference between firms with and without AI. That is exactly why such inside views are valuable, even if, as the frequent calls for independent data on AI use suggest, they only show part of the picture.
Bottom line: the entry point has to be rebuilt
Anthropic draws its own conclusions in the report: employees should share in the productivity gains, companies should use AI not just to cut costs but to create new tasks, and policymakers should strengthen retraining and income support, because protections for current employees do not offset the decline in entry-level hiring. The real task, though, lies with employers. If agents take over the routine work that newcomers used to learn on, new paths to building judgment are needed, for example as supervised reviewers and directors of agents. Companies that build those paths now will secure the experts who could be scarce ten years from now.