Anthropic cut the share of junior engineers in its technical hiring pool from twenty-seven percent down to nine percent, according to a report published by the Anthropic Institute. The creator of the Claude chatbot is the first major employer that, based on its own salary data, shows what happens to entry-level jobs when agents do the work.
The developer teams at the company found that coding and debugging tasks usually assigned to newcomers could be handled by Claude instead. Engineering and research managers subsequently stopped prioritizing lower-level technical hires. Roughly one-third of machine learning leaders and one-fifth of software development leaders admitted in an internal survey that they hired fewer people because of artificial intelligence. In the technical departments combined, the share of junior employees dropped from twenty-one to fourteen percent.
The Internal Data Behind the Hiring Shift
The study, published on October 16, was written by economists Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory. It draws on internal employment records, calendar metrics, Slack logs, and agent usage statistics collected from the beginning of 2025 through the summer of that year, alongside seventeen staff interviews. Postings for junior roles began falling sharply in December 2025, and engineering teams actively deprioritized junior hiring by April 2026. Across all technical departments, the share of junior employees dropped from twenty-one percent to fourteen percent.
Autonomous agents sustained high operational workloads during the study period. Claude registered six times as many active work hours as the human supervisors managing the system, counting only sessions initiated by employees. Code contributions per engineer arrived five times faster than they had six months prior, while lawyers reviewed more contracts and sales staff closed more deals. Employees who had never previously contributed code went from averaging two agent coding sessions per month to ninety. One sales representative built a custom dashboard in three to four hours, completing a task previously considered outside their skillset.
Human Attention as the New Operational Bottleneck
All generated output required human review, turning staff attention into the primary operational constraint. During interactive sessions, Claude systems sat idle roughly eighty percent of the time while waiting for human replies. Non-technical staff members spent more than half of their working hours simply reading text generated by the agent. One engineer described the environment as an army of Claudes demanding constant oversight, comparing the supervisory burden to childcare where simple requests arrived with extreme urgency. Written instructions given to the agents expanded from a few hundred to 8,500 over a single year. Meta-work, including writing run specifications and monitoring workflows, consumed twenty-two percent of agent sessions, up from ten percent.
Meeting schedules expanded rather than shrank. Calendar data showed median employees spent twenty percent more time in meetings compared to the previous year, while engineers without management responsibilities logged a sixteen percent increase. Staff members described face-to-face coordination as indispensable, noting that a pause from managing agents typically meant joining another meeting.
Changing Expectations for Modern Candidates
Hiring criteria shifted away from deep domain knowledge toward independence and personal judgment. Recruitment managers placed greater emphasis on how effectively candidates utilized AI tooling, with explicit notes on AI proficiency tripling for technical listings and quadrupling for non-technical roles. Employees inside Anthropic estimated that AI could execute seventy-five percent of their daily tasks by 2030, contrasting with twenty-eight percent among surveyed American workers generally.
The report itself acknowledged clear limitations. Anthropic tripled its headcount during the last year and paid portions of compensation in stock, making it an unrepresentative sample for a typical firm. The researchers noted that engineers utilizing autonomous agents were already on steep productivity curves before the transition began. Furthermore, internal testing for subpar AI code relied on limited indirect metrics. An early warning system designed to transition existing workers into new roles provided no relief for applicants who never secured an interview in the first place, leading the authors to note that job security for current staff does not help those who are never hired.
