1. What did Anthropic report about junior hiring?
On October 16, 2026, Anthropic published "How employment is changing at Anthropic," a technical report from The Anthropic Institute. It shows that junior employees fell from 21% to 14% of technical-department staff since early 2025, and that juniors made up just 9% of recent technical hires, down from 27%. Anthropic junior hiring has dropped even as the company has kept recruiting.
The authors are Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. They drew on employment and headcount records, AI usage logs, output metrics, employee surveys, Slack communication patterns, calendar data and hiring scorecards, together with 17 employee interviews. Automated classifiers built on Claude were used to label agent logs, task difficulty, meta-work and interview comments.
2. How abrupt was the shift away from junior roles?
The report describes the change as sudden rather than gradual. Junior postings fell sharply after December 2025, and the share of junior employees in technical departments declined from 21% to 14% over the period since early 2025. Among technical hires specifically, the junior share dropped from 27% to 9%.
Technical departments also became a smaller part of the whole company. Their share of total headcount slid from 61% to 56% over the past year. The report is careful to note that technical hiring still grew in absolute terms. It simply grew more slowly than hiring in sales, finance and recruiting, which expanded faster.
Non-technical staff, meanwhile, are doing more technical work. The number of coding-agent sessions run by non-technical employees rose from about two per month in August 2025 to about 90 per month in May 2026.
3. Who is writing code at Anthropic now?
The report suggests that the boundary between technical and non-technical work is blurring. The share of technical staff with no code commits in the prior 30 days fell by nearly half over the year through April 2026, which means that people in technical departments who once wrote little code are now producing more of it with agent help.
Roles that were not traditionally hands-on are changing as well. The ratio of engineers to product managers is about 17 to 1 and rising, and the number of code commits made by managers doubled. Administrative hiring, by contrast, has grown slowly, so the expansion of the workforce has been concentrated in customer-facing and commercial functions rather than in support roles.
Agents are also handling harder work than before. The share of tasks above the 90th percentile of difficulty rose from 10% to 13% after new model releases, according to the report. In practical terms, this means that the most demanding tasks, which might once have been handed to a more experienced human, are increasingly being attempted by agents.
These shifts help explain why the junior share has shrunk. Entry-level technical roles have traditionally been a place to learn by handling well-defined coding tasks under supervision, and those are precisely the tasks that agents now take on. When a manager can obtain the same output from an agent session, the case for adding a junior engineer to a team becomes weaker, particularly when the same manager is also spending more time reviewing machine output and attending meetings.
4. How many hours do AI agents work compared with people?
Agent work hours first passed human work hours in March 2026. By August, the interactive agents directed by an employee logged about six times the hours of the human directing them, with human hours assumed to be 40 per week. That is the basis for the claim that agents now outnumber staff six to one in time worked.
Output followed. Code commits per engineer reached about five times the level of six months earlier. Output rose in other functions too, including contracts handled by lawyers, invoices processed by accounts payable and deals closed by sales teams. In one example from sales, an account executive estimated that automations replaced about 20% of earlier tasks and allowed contact with roughly 30% more customers.
The report also looked at a within-person comparison of 192 engineers who adopted coding agents against 74 who did not. Code output increased after adoption, and agent-written and human-written code showed no difference in how often it was later deleted or reverted. A caution applies: engineers who adopted fully autonomous agents were already on a steeper output path before they switched.
5. Are managers hiring less because of AI?
An April 2026 survey of managers gives the clearest evidence on intent. A majority said they now hire for different skills because of AI. About one-third of machine learning managers and about one-fifth of software engineering managers said they hire less because of AI.
Hiring scorecards point the same way. Mentions of a candidate's use of AI tools were about 5% of technical and 7% of non-technical scorecard comments in 2024, and by 2026 they had roughly tripled and quadrupled respectively. For technical candidates, comments about autonomy and judgment rose most, while comments about domain expertise shrank. For non-technical candidates, customer and mission orientation grew and remarks on execution skill fell.
6. Is the extra output free of extra work?
The report argues that it is not. About 80% of interactive Claude session time is idle, because the agents are waiting for humans. For non-technical groups, reading all the output of interactive sessions would take more than half of a work week.
Meta-work, meaning building tools for agents, reviewing their output and monitoring pipelines, climbed from 10% to 22% of interactive sessions between fall 2025 and spring 2026. Standing instructions given to agents grew from a few hundred in July 2025 to about 8,500 a year later. Median employee meeting time rose about 20% year over year, and technical individual contributors' meeting time rose about 16%.
The authors say agents appeared in nearly half of active public Slack channels. Rather than displacing human communication, they added to it, and they depend on work being recorded in visible channels.
The implication for any employer is that raw commit counts overstate the gain. Reviewing and checking machine output is real labor, and a firm that ignores it may find that efficiency improvements are smaller than the headline numbers suggest.
7. What do employees say about their own productivity?
In late 2024, 44% of employees agreed that their processes made them as productive as possible, and the figure was 58% in November 2025. In April 2026 it was 75%, although the question dropped the words "as possible," so the numbers cannot be compared strictly. In an April 2026 open-ended survey, 27% reported productivity gains and 20% said their role had shifted toward managing Claudes.
Employees also expect more change. They anticipate that AI will handle 75% of their work by 2030, compared with 28% for other Americans in an August survey. Employees with at least a year of tenure were about 50% more likely to change departments in a given month in 2026 than in 2025, which suggests that roles are being redrawn as tasks move to agents.
8. Why does Anthropic's experience matter beyond one company?
Anthropic is effectively the first adopter of its own technology. Its staff had access to advanced agents earlier than most enterprises, so its workforce may preview what heavy AI adoption looks like elsewhere. The report also states that revenue per employee at Anthropic, at about five years old, is reportedly two to four times what Meta and Google achieved at the same age, based on 10-K filings for 2003 and 2009 deflated to 2025 dollars.
A fast-growing company would normally add many junior workers. That Anthropic did not, and that the report ties this partly to AI agents, makes the case notable. A fall from 27% to 9% of technical hires at a firm in expansion could make the market harder for graduates and early-career engineers if other companies follow.
The report does not claim that the same pattern will appear at every employer. It does say that the combination of rapid growth, rising revenue per employee and a shrinking junior pipeline is unusual enough to merit close study, and that the company intends to keep tracking role-level demand as AI use deepens.
9. What policy responses does the report suggest?
The authors propose stock ownership as a means of sharing productivity gains, comparing it to the postwar bargain between unions and management. They say Anthropic runs an early-warning system that tracks role-level demand and productivity against AI use, and that it prioritizes redeploying people internally.
They acknowledge trade-offs. Internal reallocation may discourage hiring from outside, though it may also steer firms toward new tasks and away from so-so automation. The report also warns that declining entry-level hiring is not covered by protections designed for incumbent workers, and it calls for retraining and income-support policy.
10. What are the caveats?
The authors flag several limits. Tool adoption is not random, since employees with large projects may adopt tools for reasons that bias the estimates. Business growth coincided with AI adoption, which complicates attribution. Code volume does not equal quality or business impact, and AI errors may be harder to spot. Some results rely on automated classifiers, some survey questions changed wording, and the interview evidence is qualitative and drawn from only 17 people.
Above all, the findings come from one company at the frontier of AI, whose employees build the technology and are probably more willing than the average worker to delegate to it. The report offers a view of one possible path, not a prediction for the whole labor market, and it identifies AI agents as a major, though not the only, likely explanation for the shift in junior hiring.