Anthropic cut junior hires to 9% — who trains the experts now?
Anthropic has published the first frontier-lab account of what AI agents do to the people who work beneath them, and the headline is that the lab is stopping halfway up the career ladder. The economics team's report, How employment is changing at Anthropic, is built from internal hiring records, agent telemetry, employee surveys, calendar data and 17 staff interviews. We walked through the core numbers yesterday — Anthropic's junior tech hiring fell to 9% as agents outworked staff. The Decoder's write-up, out today, adds two chart-level facts worth more than their space: open junior technical roles have shrunk to 6% of openings, and inside research departments agents now log 19 hours of work for every human hour, according to The Decoder. The hiring data is the news. The report's quieter claim — that Anthropic is dismantling the rung on which expertise is built, and knows it — is what deserves the depth.
What the numbers actually show
Junior share of technical headcount: 21% to 14% since early 2025. Junior share of technical hires: 27% to 9%. Junior postings have fallen sharply since December 2025 — all straight from the report. Set that against a company that tripled in size while technical departments slipped from 61% to 56% of staff, and part of the story is composition: sales, finance and recruiting scaled the way any fast-growing business would, and junior hiring in non-technical departments held roughly steady, as The Decoder notes. The report itself documents the break: in April's survey, about a third of ML managers and a fifth of software managers said they were hiring less because of AI.
The company also ran an automated read of thousands of interview scorecards filed since 2024. The biggest increase in emphasis was on candidates' autonomy and judgment; domain expertise shrank as a share of assessment. One hiring manager's explanation cuts both ways: models will stay "jagged," so the team should be "fungible across domains" — which is an argument for judgment, and also an argument that the deep vertical knowledge juniors used to accumulate is no longer what Anthropic buys. On the output side, code commits run at roughly five times the pace of six months earlier, and The Decoder's read of the report's charts puts the lawyers at 3.1 times the contracts, sales at 1.7 times the deals, recruiting at 1.4 times the hires, accounting at 1.4 times the invoices.
The constraint moved with it. Median employees now spend 20% more time in meetings than a year ago — 16% more even among technical individual contributors — and one engineer describes the job as facing "an army of Claudes, with every one asking for my attention," as the report puts it. Attention, not code, is the scarce input, and the juniors who used to absorb routine work are not being hired into the queue.
The apprenticeship hole
This is where the report stops being a productivity story. Nolan Lovett of the NATO Special Operations University, writing in Human Resource Development Review, argues that rational AI adoption by any single firm drains a shared resource: professional expertise. His two mechanisms are worth separating. First, entry-level roles are eliminated outright — the direct version, and the one Anthropic's charts measure. Second, subtler: juniors who use AI hit output targets years early, so the slow, error-laden practice that actually builds domain judgment never happens, as The Decoder summarized the paper. Either way the follow-on problem is the same: reviewing plausible-looking AI output requires exactly the depth that heavy AI use wears away — Lovett calls it a validation tether.
The timing is why this matters now rather than later. Lovett's estimate is that the damage won't surface until 2030 to 2045, because today's available experts were trained 5 to 20 years ago. The most exposed fields he names: software engineering, financial analysis, legal research. Who wins in the meantime: Anthropic, which now takes in between two and four times the revenue per employee that Meta and Google did at the same age, and senior staff whose judgment becomes the bottleneck input everyone queues for. Who loses: entrants first, then every employer downstream that needs someone qualified to sign off. The report concedes the shape of the problem — its own remedies (broad stock ownership, an early-warning system, internal redeployment) all protect people already inside the building, and none of them reopen the entry door it just closed. Outside evidence points the same way: Stanford's Erik Brynjolfsson found young workers in AI-exposed occupations down 19%, which we covered in AI's entry-level gap hits 19% — but Brynjolfsson sees no job apocalypse, and Anthropic's own bibliography cites a study of US résumés and job postings finding generative AI adoption is seniority-biased.
The case for shrugging
The report's own caveats are printed next to its findings: one company, self-reported, tool adoption chosen rather than assigned, and explosive business growth running underneath every productivity number. The composition explanation for the hiring shift is real — non-technical junior hiring held steady while sales and finance scaled. The macro picture is thinner than the lab's interior suggests: the wider economy has yet to show measurable productivity gains from AI, The Decoder points out. And Lovett's mechanism only compounds while humans remain the ones doing the end work — a point Anthropic itself underscores when it suggests programmers and call-center staff may eventually move into trades like electrician or nurse. The lab's own economic modeling makes the worst labor outcomes a fast-growth scenario rather than a forecast; we mapped that in Anthropic's 2030 model: fast growth is the bad scenario.
The shrug has a hole in it, though. Anthropic didn't just report headcount — it showed a causal design (192 engineers who adopted coding agents against 74 who didn't, with agent commits no more likely to be reverted within 90 days) and a hiring-scorecard shift toward judgment and away from expertise. Growth alone doesn't explain why interviewers stopped asking about domain depth. A lab that models the bad scenario and then lives it is not comfort; it's a dress rehearsal.
What to watch
Whether any other lab publishes the same disclosure — Anthropic has set the precedent, and the number to copy is junior share of openings, not hires, because postings lead headcount by months. Whether the early-warning system shows up next report as rehiring rather than only redeployment. And the leading indicator that predates 2030: CS enrollments and apprenticeship slots. The expertise shortage will be visible in course catalogs years before it appears in salaries — and those catalogs are being written now.
If the lab best placed to measure AI's effect on entry-level work is also the one cutting those roles, whose numbers should we trust — and who is counting the experts nobody is training? Write to us via our contact page.
