Anthropic's junior hiring collapse shows AI's first job casualty is the apprenticeship ladder

Anthropic has almost stopped hiring juniors into technical roles, and it says the reason is its own AI. In an internal study the company published, junior positions fell from 21% to 14% of technical staff since early 2025. Among new technical hires the drop is much steeper, from 27% to 9%. Open junior roles fell to 6%, with the sharpest decline starting in December 2025. Non-technical departments stayed mostly stable.

The study ties this to agent workload. According to Anthropic, its AI agents have logged more work hours than human employees since March 2026. By August the company-wide ratio was 6 to 1, and in research departments it was 19 to 1. Writing code and finding bugs go to Claude more and more often. Junior hiring fell most in the roles where Claude does the most complex work. The company also reports productivity gains in every department. Engineers shipped 5.1 times as many code commits, data scientists ran 3.1 times as many queries, and the legal team processed 3.1 times as many contracts. Sales closed 1.7 times as many deals. Recruiting and accounting each handled 1.4 times their previous volume.

These numbers need context. Anthropic itself says it is a special case: an extremely AI-heavy software company whose results don't map easily onto the broader economy. It also notes that rapid business growth and voluntary adoption of AI tools could be skewing the figures. The metrics also deserve some scepticism. An "agent hour" is not a human hour, and a commit is not a unit of value. More commits can mean more output. It can also mean more code that someone has to verify. As The Decoder notes, the wider economy has not yet shown measurable productivity gains from AI. This is a lab measuring itself with its own tools, and it should be read that way.

The more important part of the story is not the efficiency. It is the cost being passed on to others. The Decoder cites Nolan Lovett of the NATO Special Operations University, who calls this a "tragedy of the cognitive commons." When a single company replaces entry-level positions with AI, it keeps the efficiency gains. The cost is spread across everyone who will need experienced talent later. Experts are built through years of work in exactly these junior roles. Lovett argues the damage won't show until 2030 to 2045, because today's experts were trained 5 to 20 years ago. He names software engineering, financial analysis and legal research as the most exposed fields. That matches what we've seen in our sector-by-sector series. The base of the law-firm pyramid is narrowing, and in tech the routine work disappears first.

Our reading: this is the clearest signal yet of how AI will reshape work in the short term. Mass layoffs are not the first effect. A missing first rung on the ladder is. For the individual company, the decision is rational. For the system, it is irrational if nobody pays to train the next generation. There is one caveat. Lovett's warning only holds if humans still do this kind of expert work in the future. Anthropic has suggested in the past that programmers and call-centre workers may eventually need to move into jobs like electrician or nurse. The company itself does not seem to expect that.

We are not catastrophists about this. In the long run, an economy where agents do the routine knowledge work can deliver abundance and free people for work that matters to them. But the transition won't manage itself. If verification and judgment are becoming the scarce resources, someone has to keep producing people who can verify and judge. That means rethinking apprenticeship: paid training roles, juniors who orchestrate agents instead of competing with them, and perhaps shared industry funding for talent pipelines. Firms that cut the ladder today may find in ten years that there is nobody left who can check the agents' work.