Junior people now make up 9% of Anthropic’s technical hires, down from 27%, according to a report the company published on 16 October. Most employers would not put a number like that in print about themselves. Anthropic did, and the finding deserves a careful reading because two figures in it are easy to mix up.
They measure different things. Junior-level staff were 21% of the people working in Anthropic’s technical departments in early 2025 and are 14% today. That is headcount: who is on the payroll. The 27% to 9% figure is hiring: who was brought in. Headcount moves slowly, since everyone already employed stays on the books, while the mix of new hires can swing within a quarter. So the hiring number falls faster. The report’s text gives shares, not counts, and the company tripled in size over the year, so it does not say fewer juniors were hired in absolute terms.
What the report ties this to is careful. Talk of lower demand for junior technical staff began in January 2026, on the teams that leaned hardest on Claude, because the model could take over bug-finding and code-writing. One manager said a new model “basically overnight changed what we were looking for.” In an April survey, about a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI. The authors call the link between Claude use and slower junior hiring descriptive, and part of the shift is simply a company adding salespeople, finance staff, and recruiters.
The productivity evidence is stronger than a company-wide snapshot. The researchers followed 192 engineers as they adopted coding agents and compared them with 74 who did not. Code output rose sharply after adoption, and code written by an engineer’s agent was no more likely to be removed or reverted than the same engineer’s own. The authors flag the weak spot themselves: adoption is not random, and the engineers who went fully autonomous had already been producing more before they made the switch. By August, agents working in employee-directed sessions logged six times the hours of the humans directing them.
The cost shows up as attention. Converted at a typical reading speed, the output of non-technical employees’ Claude sessions alone would fill more than half a work week (an estimate from token counts, not a measurement of anyone reading). One engineer described “an army of Claudes, with every one asking for my attention.” In the same April survey, 27% of open-ended answers described productivity gains and 20% described the job shifting from doing the work to managing Claude. That 27% has nothing to do with the hiring figure.
Two limits apply, and they pull in opposite directions. This is a company studying itself on a subject where it benefits from a story of powerful AI, and the headline productivity numbers come from its own telemetry. It is also one firm, which the report describes as unusually thoroughly built around AI, so it is a leading indicator at best and says nothing reliable about software employment or the wider economy. Against that, the authors write that none of the company’s own measures, from stock ownership to moving staff between teams, “directly address the slowdown in junior hires,” because protections for current employees “will not benefit those not hired to begin with.”
The report does not raise the following point, so it is our framing. The entry-level rung is where most people learn to become senior. A firm that stops hiring beginners is borrowing against its own future supply of experienced people, and it is relying on being able to buy them elsewhere. If every firm that adopts agents this thoroughly does the same, that supply has to come from somewhere. The report quantifies how fast one company stopped hiring its beginners; it does not say who trains the next set of seniors, and nothing in it claims to.
Based on Anthropic’s research report “How employment is changing at Anthropic,” published 16 October 2026.
