The Pulse
Anthropic Agents Outworked Employees by Six to One in August
Anthropic says its AI agents surpassed employees’ work hours in March 2026 and logged six times as many hours as their human handlers by August. The company’s study also tracks shifts in junior technical hiring, employee output and the grow
AI.info Team · Written with AI ·
By August, Anthropic’s AI agents were logging six hours of work for every hour logged by the employees directing them. The company says total agent work hours first passed human work hours in March 2026, a striking measure of how far it has reorganized its own operations around AI.
The figures come from an Anthropic Institute study published October 16, 2026, drawing on company employment records, agent telemetry, output measures, surveys and interviews with 17 employees. They describe one AI lab, not a typical employer: Anthropic builds software and develops the very systems its staff use. The authors present its experience as a possible early signal of workplace change, not proof that other companies will follow the same path.
What Anthropic counts as agent work
The March comparison combines autonomous and interactive agent work. The August six-to-one figure is narrower: it counts employee-directed interactive sessions and pools the work of subagents. Anthropic defines agent working time as runtime spent on inference and active tool use, excluding waits longer than 30 minutes. For the human comparison, the researchers estimate employee hours from headcount, assuming a 40-hour workweek.
That distinction matters. The comparison is a measure of recorded agent runtime against an estimate of human working hours; it is not a count of employees replaced, nor a like-for-like measure of useful output. The study says much of the activity involves software engineering, including tests, maintenance and work across large codebases. Agents also help with calendars, messages, meeting briefs and other routine tasks.
Junior technical hiring falls as output rises
Anthropic’s hiring mix shifted during the same period. From early 2025 to the study’s reported 2026 comparison, junior employees’ share of technical departments fell from 21% to 14%; the share of new technical hires at junior levels dropped from 27% to 9%. Technical departments’ share of the overall workforce also declined, from 61% to 56%, even as the company continued adding technical staff in absolute numbers.
The researchers report higher measured output across several functions. Code commits per engineer reached five times their pace six months earlier, while lawyers processed more contracts and sales staff closed more deals. But the study cautions against treating code volume as a complete productivity measure. “More code does not necessarily mean better code or business impact,” the authors write.
“More code does not necessarily mean better code or business impact.”
Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, The Anthropic Institute
To examine whether adoption tracked with changes in output, the study followed 192 engineers who adopted coding agents and compared them with 74 who did not adopt during the period. The authors found that code added by adopting engineers rose after they began using the tools. They also warn that adoption was not random: workers taking on major projects may have been more likely to adopt agents, which could affect the results. Anthropic says it found no difference in several measures of how often agent-written code was later removed or reverted, while acknowledging that those checks cannot capture every dimension of quality.
More agent output means more work for people
The study’s picture is not one of human oversight disappearing. Anthropic reports that interactive agent sessions spend about 80% of their wall-clock time waiting for human input. In April, 20% of employees who answered an open-ended survey described a shift from doing work themselves to managing Claude agents. The share of interactive sessions devoted to “meta-work”—such as writing agent plans, building dashboards or monitoring pipelines—rose from 10% in fall 2025 to 22% in spring 2026.
For non-technical employees, the volume of agent output also created a review burden: the authors estimate that reading all the text produced in interactive sessions would take more than half a workweek. The account complicates a simple substitution story. Agents can extend the amount of work a team produces, while shifting human effort toward directing, checking and organizing that work.
Anthropic’s records also show more movement between teams: employees who had worked at the company for at least a year were 50% more likely to change departments in 2026 than in 2025. The report leaves an important question open: whether employers outside an AI lab, with different tasks and less concentrated technical expertise, can achieve similar output gains—and what happens to entry-level hiring as they try.