Anthropic's junior technical hiring falls from 27% to 9%
The company's own report says agents now log six times the hours its people do, and sit idle 80% of the time waiting for a human to reply. The bottleneck has moved, and it is you.
Anthropic has published a report on what agents have done to its own workforce. It draws on internal headcount data, agent session telemetry, calendar and Slack patterns, hiring scorecards and 17 employee interviews.
The headline number is a hiring one. The share of new technical hires who were junior fell from 27% to 9%, according to the report. The junior share of staff in technical departments fell from 21% to 14% over the same period.
The company dates the shift to January 2026, starting in the teams that used Claude most. Writing code and finding bugs went to agents first.
By April, the report says, engineering and research had deemphasised lower-band technical hiring. About a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI.
Anthropic concedes its worker protections do not cover this. Redeployment helps people already on payroll. It does nothing for the engineer who never got the offer.
Headcount still tripled in the past year. Technical departments shrank as a share of the company, from 61% to 56%, while sales, finance and recruiting grew faster.
The second number founders should sit with is 80%. Interactive Claude sessions spend roughly four-fifths of their time idle, waiting for a person to respond.
Reading load explains much of it. Anthropic converted agent output into reading time at normal human speed and found that even non-technical teams would need more than half a working week to read everything their agents produce.
Total agent hours passed total human hours in March 2026, on the company's figures. By August, counting only agents employees directed themselves, agents worked six times the hours of their humans.
The management overhead is growing with it. Written standing instructions for agents went from a few hundred in July 2025 to 8,500 a year later, and agent management rose from 10% to 22% of the average employee's interactive sessions.
Meetings went up too. The median employee spends 20% more time in meetings than a year ago. Cross-team decisions still get made by humans, and agents cannot read a decision taken in an unrecorded meeting or a DM, so they keep working from stale context.
What interviewers look for has moved with the work. Anthropic ran its hiring scorecards since 2024 through an automated coder and found mentions of a candidate's track record with AI tools tripled for technical roles and quadrupled for non-technical ones. Comments on autonomy and judgment grew most; domain expertise shrank as a share of the assessment.
The authors flag a causality caveat on their productivity figures: engineers who went fully autonomous were already on a steep output trajectory beforehand. The numbers are directional.
For a small team, the practical reading is cheap. Hiring a junior to write code solves a problem agents have largely absorbed; hiring someone who can direct agents and clear review backlog solves the one Anthropic measured at 80%. The constraint on a two-person company running agents is not agent capacity. It is how fast the two of you can read.