Anthropic says its junior technical hiring share fell from 27% to 9%

The company whose own agents are being used to write most of its code has just published what that did to its hiring.

Anthropic released a report on 16 October called "How employment is changing at Anthropic". The authors, Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory at the Anthropic Institute, drew on headcount records, agent session logs, code commits, three employee surveys, calendar data, Slack network analysis and 17 interviews. The main sampling period is October 2025 to August 2026.

From 27 percent of new technical hires to 9 percent

Anthropic says the junior share of its technical workforce fell from 21 percent to 14 percent between early 2025 and August 2026. The junior share of new technical hires in the same period fell from 27 percent to 9 percent. The share of the whole company working in technical roles also dropped, from 61 percent to 56 percent over one year.

A survey of hiring managers in April 2026 recorded that a majority of managers said they were hiring for a different set of skills because of AI. The company says around a third of machine learning managers and a fifth of software engineering managers reported hiring less because of AI.

Six agent hours for every employee hour

By August 2026, Anthropic says interactive agent hours were running six to one against employee hours. A separate measure of agent output showed that reading all of it would take an employee about half a work week on its own.

The report also says code commits per engineer were five times higher than six months earlier, and that engineers who had adopted coding agents (192 of them in the sample) shipped more than the ones who had not (74). Anthropic says the adopters chose themselves rather than being assigned, and that extra commits do not have to mean better code.

Meta-work, by which the authors mean the share of Claude sessions people use to plan, review or coordinate other work rather than do a task directly, rose from 10 percent in autumn 2025 to 22 percent in spring 2026. Meeting time for technical individual contributors rose 16 percent year on year.

Why the authors say the figures are not a clean test

The paper says tool adoption inside Anthropic is not random, so the productivity figures are not a clean test of what agents do for an average team. It also says many of the practices behind these numbers, including unusually high internal AI use, cannot be copied across the economy. The agent-to-employee ratio is weighted by the fact that an agent session is idle for about 80 percent of its length waiting on a person.

The report is Anthropic's own account of its own data, built on internal records and surveys, and no outside auditor is cited. Anthropic sells Claude, so the productivity figures sit under the usual caveat that a vendor is reporting on the value of its own product. The hiring numbers are a different kind of fact: they are what one company that uses coding agents heavily decided to do with its headcount over 20 months.

Source

How employment is changing at Anthropic, by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory.