Anthropic's junior share of technical hires fell from 27% to 9%. Its interviewers now write about judgement

Anthropic published a study of its own workforce on 16 October, setting its staff's use of Claude agents beside its hiring records. Gene Dai's analysis of that study, published on 18 October, pulls out the figure that matters for anyone planning their next role: the junior share of Anthropic's technical hires fell from 27% to 9% since early 2025. Anthropic says it has tripled in size in the last year, and technical hiring kept growing in absolute terms. The study does not count jobs lost to Claude, and Anthropic sells the tools it uses, so its numbers do not transfer cleanly to any other employer. What does transfer is a record of what the people doing the hiring started writing down about candidates.

What the scorecards now reward

Anthropic analysed its interview scorecards from 2024 to 2026. Mentions of candidates' creative use of AI tools rose substantially. For technical candidates, comments on autonomy and judgement rose as well. Discussion of narrow domain expertise became less prominent.

The managers said the same thing when asked directly. By April 2026, around one-third of machine-learning managers and one-fifth of software-engineering managers said they were hiring less because of AI. A majority said they were hiring for different skills.

The obvious reading is that juniors are out and seniors are in. The scorecards say something sharper. Years of experience are a proxy. What interviewers are recording is whether a person can be handed an ambiguous problem and a set of agents, and be trusted to come back with a sound decision. Seniority correlates with that. It does not guarantee it. A less experienced candidate who can show it is being graded on the same rubric.

The job inside the job

The study also shows what the work turned into once agents arrived. By August, employees' interactive agents logged six hours of active runtime for every assumed hour of human work. Dai points out that this compares compute time with a standard 40-hour week, so it measures the scale of agent use, not payroll hours saved. In multi-turn sessions, agents spent about 80% of elapsed time idle, waiting for a human.

The share of interactive session time spent on meta-work rose from 10% in autumn 2025 to 22% in spring 2026. Anthropic's meta-work category covers managing agents, building Claude skills and dashboards, and documenting how work is done; the company says much of the rise was agent management. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. In an April survey, 20% of employees described a move from doing tasks to managing Claudes. The median employee spent 20% more time in meetings than a year earlier.

That is the work the scorecards screen for. Someone decides what to run, writes the instructions, judges the output and stops the run that is going wrong.

AI does the routine work. You do the thinking. Anthropic's records show what that line looks like in a calendar.

Movement goes to people already trusted

One more figure from the study. Staff who had been at Anthropic at least a year were 50% more likely to move between departments each month in 2026 than in 2025. A product-content team whose routine copy went to agents moved into prompting and model evaluation.

Most employers are not Anthropic. Workday's October 2026 Global Workforce Report, built partly on de-identified data from its customers, found that internal moves fell at 57% of employers year over year and promotion rates stayed essentially flat. Its job-requisition data shows mentions of management and leadership skills fell 7%, and training skills 13%, between September 2025 and July 2026. Employees who tried to move internally most often cited paused hiring, managers who did not support the move and selection processes that did not feel fair.

The contrast is the signal. At a company running six agent hours for every human hour, tenured staff moved between departments more often than a year earlier. Across Workday's customer base, the internal route is narrowing, and the obstacles employees name are managers and selection processes. A documented case for your move is the part of that you control.

Your Next Move

Keep a judgement log. For the next four weeks, record every time you reject, correct or stop an AI output. Three lines per entry: what the tool produced, what was wrong with it, what you decided instead. Anthropic's interviewers are now writing about autonomy and judgement. This log is the evidence for both, and it is the thing most people cannot produce when a review or an interview asks for it.

Own one agent workflow end to end. Pick a recurring task on your team. Write the standing instructions, define what an accepted output looks like and put your name on the review. Meta-work rose from 10% to 22% of interactive session time at Anthropic between autumn 2025 and spring 2026, and Anthropic attributes much of that rise to agent management. Be the named owner of that work before it is handed out by default.

Ask for the rotation before the vacancy is posted. If a team near yours is moving into evaluation, agent oversight or AI-enabled process work, raise it with your manager this month, log in hand. An unsupportive manager is one of the blockers Workday's respondents named most often. A documented record of your decisions is the strongest answer to it.