Anthropic says AI agent hours at the company now outnumber employee hours six to one

Anthropic’s economics team published a case study of its own workforce, arguing that a company that has built its workflows around AI may preview changes elsewhere. The authors are Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. The headline finding is that “agent hours worked outnumber human hours 6 to 1”. The report says agent hours passed human hours in March 2026, and that by August interactive agent hours were six times employee hours, measured as runtime with waits over 30 minutes excluded, against an assumed 40-hour week per employee.

The report draws on internal records for employment, hiring decisions, AI use and output, plus employee surveys, Slack and calendar data, and 17 interviews. It says code commits per engineer were five times their level six months earlier. In a staggered-adoption comparison of 192 engineers who took up coding agents with 74 who did not, adopters’ output rose sharply afterwards. Anthropic said it found no difference in the rate at which agent-written and human-written code was later removed or reverted.

On hiring, the report says the technical share of staff fell from 61% to 56% over the past year and the junior share of technical employees from 21% to 14% since early 2025. Junior staff made up 9% of technical hires, down from 27%. About a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI. Roles have shifted too: coding sessions by non-technical staff rose from 2 a month in August 2025 to 90 in May 2026, and the share of interactive sessions spent on supervising agents and building tools (“meta-work”) rose from 10% to 22%. Human attention became the constraint: about 80% of interactive session time was spent waiting for a person.

The authors listed their own limits. Employees who adopt agents may differ from those who do not, and adopters were already on a rising output trend. Rapid company growth, which tripled headcount, could drive some results. More code is not necessarily more value, a survey question changed wording between waves, several measures rely on Claude-based classifiers, and the junior-hiring analysis is descriptive, not causal. They also said the company’s own measures do not directly address the slowdown in junior hiring, and that government policy on retraining and income support is needed.

The study is single-company and self-reported by a firm that sells the tools it describes. It appeared within two weeks of Kroger and HubSpot announced job cuts they framed around AI, and supplies company-level data on how junior roles change when coding agents are widely used, though from one firm.