What the report says
In a report published on Thursday, Anthropic states that adoption of its Claude-based agents has changed the organization of internal work: staff reportedly spend about 22% of their time managing agents, more than double the figure a year earlier, and since May 2026 the agents have reportedly logged more work hours than employees, according to an analysis of chat summaries cited by the Guardian. The company also claims to produce more code per person, close more sales per employee, and handle more contracts.
The most significant figure concerns employment: since March 2025, shortly after the launch of the agentic coding tool Claude Code, entry-level technical hiring has reportedly fallen from nearly 30% to 9%, and overall headcount from 21% to 14%. Anthropic attributes these changes to the use of AI and describes its own employees as a "thin layer" above a growing number of agents. The report envisions a future scenario of startups led by a single person directing swarms of agents.
It is appropriate to distinguish facts from statements: the figures are provided by the company itself and are based on an internal analysis, not on independent verification. The same CEO, Dario Amodei, the Guardian recalls, had predicted last year that AI would eliminate half of entry-level office jobs and write 90% of code within six months; neither prediction has come true.
Criticisms of method and substance
Several researchers question both the validity of the data and the motives behind the publication. Morgan Frank, of Case Western Reserve University, has circulated a paper in peer review according to which labor markets exposed to AI deteriorated long before the arrival of ChatGPT in November 2022 and of Claude a few months later: unemployment claims would show a worsening already from the beginning of 2022, without the launch of ChatGPT corresponding to a significant increase in job losses. Frank cites among possible factors the cutback in pandemic-era hiring and "AI-washing," that is, attributing to AI layoffs that would have occurred anyway.
Julián Posada, a social scientist at Yale and author of a book on platform extractivism, observes that the report is written from the perspective of a company that wants to sell shares and attract investors, and that an independent assessment by external auditors is missing; the figures and graphs, according to Posada, tell only part of the story. Dylan Baker, of the Distributed AI Research Institute, calls the strategy of shifting part of software development onto AI while reducing junior hiring short-sighted, and considers the description of human attention as a "bottleneck" bizarre: software development, he argues, is a social and interpersonal process. An anonymous engineer cited in the report speaks of "an army of Claude" demanding attention, and Baker doubts that well-designed products can emerge from such an environment.
The financial context and what remains to be verified
The report arrives at a delicate moment: according to the Guardian, Anthropic could debut on the stock market as early as November, one day after announcing a commitment to allocate 5% of its shares to the "public good" without providing details. The draft S-1 filing reportedly reveals a net loss of $80 billion in the first half of the year and a junk-grade credit rating obtained in June; with revenue growth slowing, investors now estimate a valuation of $1.5 trillion, down from previous estimates of $2 trillion or more. The cloud partner CoreWeave, the newspaper adds, is reportedly overwhelmed by AI-related debt, with its stock sharply down.
Anthropic did not respond to the Guardian's request for comment. The internal data on employment and productivity, the causal link between agent adoption and reduced hiring, and the actual impact on employees and on the quality of work therefore remain to be independently verified. The report offers a snapshot of a single company, but general conclusions about the future of work in the technology sector, such as those about eliminating half of office jobs, are not supported by the available facts.
