Anthropic report: AI agents now work 6 times more hours than its staff

On October 16, Anthropic released an internal study of its own workforce. Its central finding: by August, employee-directed interactive agents were logging six hours of work for every one hour worked by staff, while junior technical roles shrank as a share of hiring.

What the report found

The authors, Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, drew on internal data covering employment, AI use, output metrics, surveys, Slack patterns, calendars and hiring records, plus 17 employee interviews. They present Anthropic as a possible leading indicator for the wider economy.

The sharpest change is in junior technical hiring. According to the report, junior staff made up 21% of technical departments in early 2025 and 14% now. Their share of technical hires fell faster, from 27% to 9%. Talk of weaker demand for junior technical staff began in January 2026, and by April engineering and research teams had deemphasized lower-band hiring. In an April survey, a majority of managers said AI had changed the skills they hired for. Around a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI.

The company has still grown quickly, tripling in size over the past year. Even so, the technical share of employees slipped from 61% to 56%, because hiring in sales, finance and recruiting grew faster. Anthropic also reports revenue per employee two to four times what Meta and Google achieved at the height of their own growth spurts. It describes itself as “clerical-light,” with agents, automated review and outsourced ticket service absorbing administrative work that usually grows with headcount.

Output up, attention as the bottleneck

Agent work hours passed human work hours for the first time in March 2026, counting both autonomous and interactive agents. Limiting the count to employee-directed interactive sessions, agents had worked six times as many hours as their human handlers by August. Humans are assumed to work 40 hours a week in that comparison.

Output rose across functions. Code commits per engineer reached five times their level of six months earlier. Lawyers processed more contracts, accounts payable cleared more invoices and salespeople closed more deals. Anthropic notes that business growth could explain part of this, so it compared 192 engineers who adopted coding agents against 74 who did not. Code added rose sharply after adoption. The company also acknowledges that adoption is not random, since people about to start big projects may tool up first. Comparing agent-written and human-written code by the same employee, month and directory, it found no difference in the rates at which lines were later removed or reverted.

That makes human attention the constraint. The report estimates that even for non-technical groups, reading everything their Claude agents produce in interactive sessions would take more than half a work week. One engineer described facing “an army of Claudes, with every one asking for my attention.”

Employees are responding with what the authors call meta-work: building skills, specs, dashboards and monitors for agents. It rose from 10% of the average employee’s interactive sessions in fall 2025 to 22% in spring 2026. The number of standing written instructions for agents climbed from a few hundred in July 2025 to roughly 8,500 a year later. The median employee now spends 20% more of their time in meetings than a year earlier, and agents appear in nearly half of active public Slack channels.

Hiring criteria have shifted too. Based on interviewer scorecards, mentions of candidates’ use of AI tools roughly tripled for technical roles and quadrupled for non-technical roles between 2024 and 2026. Comments on autonomy and judgment grew most for technical candidates, while domain expertise took a smaller share of the discussion.

Why it matters

The report’s own data cannot show whether Anthropic’s experience carries over to other employers. It comes from one company that tripled in size in a year, written by its own economists, and the figures are given as shares and ratios. The 6-to-1 agent-hours comparison rests on an assumed 40-hour human week, and the authors concede that adoption of coding agents was not random. Outside readers cannot check any of it, because the underlying records are internal.

Independent evidence on junior hiring points the same way, but it is narrower. In a Stanford Digital Economy Lab working paper built on ADP payroll data, the authors report that workers aged 22 to 25 in AI-exposed occupations saw 16% relative employment declines in the November 2025 version, while employment for experienced workers in the same jobs stayed stable. They say the declines are concentrated in occupations where AI is more likely to automate than augment labor, and describe the findings as early evidence consistent with an entry-level effect, not as proof.

Other research disagrees. The Register reported that a Danish study found generative AI had no real impact on jobs or wages, and that an NBER paper found limited impact, with lower demand for some roles balanced by higher productivity in others. The Stanford authors argue those studies lacked sufficiently granular or timely data. Anthropic’s report cannot settle that dispute: a firm that builds the tools and leads the industry in using them is the least typical test case. Whether its 14% junior share is an early signal for other companies, or a feature of this one, will take external payroll data to show.

What’s next

Anthropic says it is standing up an early warning system that tracks role-level demand and productivity metrics against AI capabilities and use. The aim is to flag jobs facing disruption early and to move affected employees into other roles inside the company. The authors point to widespread stock ownership and internal reallocation as two features other employers could learn from.