Picture an engineer at nine on a Tuesday morning, a person invented for this article. She has eleven agent sessions open. Four are finished and waiting for her to read what they wrote. Two have stopped on a question that only she can answer. One has been idle since Friday because it asked, politely, whether it may delete a directory. Each of them signals with the same urgency. She has not written a line of code since breakfast, and she is the slowest component in the room.

The scene is made up, but its mechanics come from a report that Anthropic’s economists published on 16 October. Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory used the company’s own records on employment, AI use, output, calendars and hiring decisions, plus 17 employee interviews, to ask what AI has done to work inside the company that builds it. Their answer has two halves that sit uneasily together. Output per person has risen sharply. The share of junior technical people the company brings in has fallen even faster.

The junior share of hires fell from 27% to 9%

According to the report, the share of junior-level employees in Anthropic’s technical departments slipped from 21% to 14% since early 2025. The share of junior technical hires dropped from 27% to 9%. Talk of weaker demand for junior technical staff began in January 2026 on the teams that leaned hardest on Claude, which found that writing code and finding bugs could increasingly be handed to it. One interviewee said that a model’s uplift for performance engineers “basically overnight changed what we were looking for.” By April, engineering and research teams had deemphasized hiring for lower-band technical positions. In a survey that month, around a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI.

The company is not shrinking. It has tripled in size in a year, and the authors say it takes in two to four times as much revenue per employee as Meta and Google did at the same age of five years, a comparison built from Anthropic’s internal records and the two rivals’ annual filings. The technical share of all employees fell from 61% to 56%, because hiring in sales, finance and recruiting outpaced it. The authors do not claim that every shift is caused by AI. Part of it, they write, is simply a growing business adding business functions.

Agents worked six hours for every human hour

The telemetry behind the productivity claim is as stark as the hiring numbers. In March 2026, combined autonomous and interactive agent hours passed human work hours for the first time. Counting only agents that individual employees direct, the ratio had reached six agent hours to one human hour by August. Human hours there are an assumption: headcount times forty hours a week.

Output followed. Code commits per engineer ran at five times their pace of six months earlier. Lawyers processed more contracts, accounts-payable staff more invoices, salespeople closed more deals. The authors know that a company selling a strong model will close deals more easily, so they compared people in the same month: 192 engineers who adopted coding agents against 74 who did not. Code added per engineer rose sharply after adoption. They also compared code written by a person with code written by that person’s own agent, in the same month and directory, and found no difference in how often lines were later removed or reverted. They add two cautions. Adoption was not random, since people on the cusp of a big project may be the ones who tool up, and those who moved to fully autonomous agents were already ramping up their output beforehand. And these are proxies for quality, not a verdict on it.

Human attention is the bottleneck

The most vivid finding concerns what the speed does to the people. In interactive sessions, agents spend about 80% of the session waiting for a person. Converting agent output into words at an average reading speed, the authors estimate that even non-technical employees would need over half a work week to read everything their Claude sessions produce. One engineer described “an army of Claudes, with every one asking for my attention.” Another likened it to babysitting, with every child’s concern voiced at the same volume.

People respond by building tools. Standing written instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. The share of employees’ interactive sessions spent on what the authors call meta-work, such as writing specs for multi-agent runs, setting one agent to review another and monitoring pipelines, rose from 10% in autumn 2025 to 22% in spring 2026. Hiring changed with it. In scorecards from interviews, mentions of creative use of AI tools tripled for technical roles and quadrupled for non-technical ones against 2024, and the biggest rise for technical candidates was in comments about autonomy and judgment, while talk of domain expertise shrank.

Human contact did not disappear. The median employee spends 20% more time in meetings than a year ago, and among technical individual contributors the rise is 16%, though growth of the company explains part of it.

What the company is doing, and what it leaves open

The authors list three features of Anthropic’s approach that they think matter: broad employee stock ownership, an early-warning system that tracks jobs at risk and a priority on moving people into other roles. Employees who had been at the company for at least a year were 50% more likely to switch departments in any month of 2026 than in 2025. The authors also name a cost: commitments to reallocate current staff could make an employer hesitant to take a chance on new hires.

Then they say the uncomfortable part themselves. None of these actions, they write, directly addresses the slowdown in junior hiring, and protections for incumbents “will not benefit those not hired to begin with.” The report is also one company’s account of itself, written by its own economists. It has no outside comparison group, and it concerns a firm with rising valuation and unusually early access to its own tools.

Outside evidence points the same way, at a coarser grain. A Stanford Digital Economy Lab paper revised on 12 August, using ADP payroll data on millions of American workers through June 2026, finds no evidence of widespread job displacement. It does find employment of workers aged 22 to 25 in AI-exposed occupations 19% below the level it would have reached had it kept pace with less-exposed peers, with the gap working mainly through reduced hiring and not through separations. Anthropic’s own December 2025 study of its engineers had flagged the worry from the other side: some feared that when producing output is easy, learning gets harder.

The seat at the table

Go back to the imagined engineer and her eleven sessions. She is not redundant. The report suggests that she is the scarce thing, and that the company is spending its effort making her time go further. What the report cannot tell her is where the next person like her comes from, if the entry rung that once produced engineers with her judgment is the one being hired for least. The economists end by calling for retraining and income support from government alongside what employers do. Her desk in the story will keep filling up faster than she can clear it. Whether anyone is trained to sit beside her is a question the data leave open.

References

  • Brynjolfsson, E., Chandar, B., & Chen, R. (2026, revised 12 August). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu
  • Huang, S., Seethor, B., Durmus, E., Handa, K., McCain, M., Stern, M., & Ganguli, D. (2025). How AI is transforming work at Anthropic. Anthropic. anthropic.com
  • Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026). How employment is changing at Anthropic. The Anthropic Institute. anthropic.com