The company that sells the automation just opened its own payroll and showed everyone the future — the door for beginners is closing, and it is closing fastest in the one place beginners used to become experts. Anthropic’s own internal study reports that juniors’ share of its new technical hires fell from 27 percent to 9 percent, and the people who will pay for that gap have not been hired yet.
This is not a layoff story. Nobody at Anthropic was walked out of the building. The company grew. Revenue in the third quarter reached $16.8 billion. It kept adding technical staff in absolute numbers, and it simply stopped adding the inexperienced. Among all employees in technical departments, the junior share slid from 21 percent to 14 percent since early 2025. Postings for entry-level technical roles thinned out sharply starting in December 2025. The company put the number on its own website in a report called “How employment is changing at Anthropic,” built from internal staffing, usage, output and hiring records plus interviews with seventeen employees.
What changed is what the agents can do. Claude now writes code and hunts bugs — the two tasks that used to constitute a junior engineer’s entire education. In March, the hours logged by employee-directed agents passed the hours logged by the employees themselves. By August the company-wide ratio was six to one. Inside research departments it was nineteen to one.
The agents do not get tired. They also do not get promoted, which is the part nobody has priced in yet.
Peter McCrory — Eighteen Months Ahead of the Economy Is Still Your Future
Peter McCrory, an economist and one of the report’s authors, supplied the line that matters more than any chart in it. He estimates Anthropic runs roughly a year and a half ahead of the broader economy in how deeply it has absorbed AI into daily work. That is not a boast about the company. It is a warning with a timestamp on it.
If the lab is the leading edge, what it is living through now is what a hospital, an insurer, a law firm and a freight broker will be arguing about in 2028. And the lab is living through a productivity surge that makes cutting the bottom rung look like discipline rather than vandalism. Engineers ship about 5.1 times as many code commits as they did six months earlier. Data scientists run 3.1 times as many queries. Legal handles 3.1 times the contracts. Sales closes 1.7 times the deals, accounting clears 1.4 times the invoices, and recruiters hire 1.4 times as many people.
Growth used to mean more desks. Now it means more agents and the same number of desks.
Nolan Lovett — Somebody Has to Be the Junior, or Nobody Becomes the Senior
There is a name for the trap, and it comes from outside the industry. Nolan Lovett, a researcher at the NATO Special Operations University, calls it a tragedy of the cognitive commons: any single employer captures the whole benefit of replacing its entry-level roles, while the cost — the disappearance of people who know how to do the work — gets spread across everyone who will need them later.
His timeline is the uncomfortable part. Lovett argues the damage will not surface until roughly 2030 to 2045, because the senior people running things today were trained five to twenty years ago, back when the first job still existed. Software engineering, financial analysis and legal research are his most exposed fields. Those are also, not coincidentally, the three areas where the lab’s own data shows the steepest junior decline.
Put a face on it. A twenty-three-year-old with a computer science degree and a portfolio of side projects is now competing for a seat that a supervised agent occupies for the price of an API call. She cannot become the engineer who reviews the agent’s output by reading about reviewing it. The path was always the same: you did the boring work badly, someone senior corrected you, and after four years you were the person doing the correcting. Remove the boring work and you have removed the correction.
You cannot supervise a thing you were never allowed to learn.
The Managers — A Third of Them Admitted It in Writing
None of this is inference from a trend line. In an internal survey run in April, about a third of machine-learning managers and a fifth of software-engineering managers said they were hiring fewer people specifically because of AI. That is not an analyst guessing at motives. That is the people holding the req sheets saying the quiet part into a form field.
The shape of the work that remains is strange. Interactive Claude sessions spend roughly 80 percent of their time waiting on a human. Staff outside engineering would need more than half a working week to read everything their agents produce. Meetings are up about 20 percent year over year. The report’s own summary of the mood inside the building is that employees have come to describe their jobs as managing teams of Claudes.
The job did not vanish. It moved up a level. The elevator simply stopped opening at the ground floor.
Erik Brynjolfsson — The Outside Evidence Agrees, Quietly
The lab’s figures would be easy to wave off as self-serving if they stood alone. They do not entirely. Economists at Stanford led by Erik Brynjolfsson have found that firms hire significantly fewer young workers in occupations exposed to AI, software development and customer service among them. A Danish study found no such gap between companies using AI and companies that do not, and Bharat Chandar of the Stanford Digital Economy Lab has cautioned that AI’s effect on overall employment still looks modest for now.
Both caveats are honest, and both describe the present tense. The Anthropic report is not a forecast. It is a receipt from an employer that has already run the experiment everyone else is still piloting. The disagreement is not about direction. It is about how fast the rest of the economy catches up to a lab that lives inside its own product.
The Bill — Paid Later, by Somebody Who Is Currently Nineteen
Anthropic deserves a sliver of credit for publishing this at all. It could have buried the junior-hire collapse in a footnote and nobody outside the building would have known. Instead it put the chart on the internet, conceded that the sample is one unusually AI-saturated company, and let the rest of us read it. That is more candor than a decade of corporate AI surveys has trained anyone to expect.
Candor is not a correction, though. The economics of the next five years point one way: the cheapest line on an engineering budget is the person who has not learned anything yet, and the most expensive line ten years out is that same person’s absence. Firms cutting the entry level to protect a margin are borrowing against an expertise base they did not build and will not be able to buy — because there will be nobody left to buy it from.
The fix is not sentimental. It is an apprenticeship the company actually pays for, junior roles booked as capital rather than overhead, and a real answer to the question every twenty-two-year-old is now asking a chatbot because no manager will take the meeting. Anthropic just handed every HR department in the country a number that makes the opposite choice look rational.
The agents put in six hours for every one their supervisors put in. Somebody eventually has to be the supervisor. That person is in college right now, sending applications to a posting that closed in December.
Sources: Anthropic, The Decoder, Trending Topics.