Anthropic Study Finds AI Agents Create a Reading Tax on Human Attention
A new Anthropic study shows AI agents now log six hours of work for every human hour, forcing employees to spend more than half their week just reading and reviewing output. The findings highlight a growing design challenge: keeping humans meaningfully in the loop as agents take on more work.
Employees at Anthropic are now directing AI agents that log six hours of work for every hour their human counterparts put in, according to a study published October 16 by four of the company's researchers. The finding, from a review of how Anthropic's own agents have changed internal work, suggests that the biggest constraint on AI-driven productivity may not be computing power or model capability but the limited attention of the people who must review what the agents produce.
The study, titled «How employment is changing at Anthropic,» found that agent sessions spent roughly 80% of their time idle, waiting for a person to respond. Even for employees who do not write code, reading everything their agents generated would consume more than half of a standard workweek. The researchers describe this as a «reading tax» that every team deploying agents will eventually pay. Anthropic, which runs its own operations on the models it builds, offers an early look at where other companies may be heading as agentic tools spread.
The numbers point to a deeper shift in the nature of work. Rather than turning employees into makers of products, agents are turning them into managers of output, responsible for judging what deserves attention and what can be safely ignored. That raises questions about who truly holds agency in a human-in-the-loop system. As the study notes, a person who cannot read what an agent did remains in the loop only on paper while the agent makes the calls.
The problem is not new. Herbert Simon warned in 1971 that «a wealth of information creates a poverty of attention,» with most of the cost falling on the receiver. Research by Gloria Mark at UC Irvine found that people stayed on a single screen for an average of 47 seconds between 2016 and 2020, down from two and a half minutes in a 2004 study. Each additional agent acts as a new sender in an inbox that was already overflowing.
The study's authors propose several design responses. Teams should treat reading time as a budget, much like page load time, measuring how long it takes to review agent output and flagging workflows whose weekly output exceeds the hours available to any person. Agent requests should be ranked by severity, with routine items kept in a digest and only decisions requiring human judgment brought to the center of attention. The comparison to hospital alarm systems is instructive: a 2013 Joint Commission alert estimated that 85% to 99% of medical device alarms do not require clinical intervention, and its database logged 80 alarm-related deaths between 2009 and 2012 as staff tuned out the noise.
At Anthropic, agents mostly post in channels set aside for human-agent collaboration, which keeps ordinary conversation threads readable. The study also found that agent-written code was removed or reverted at the same rates as code written by hand by the same engineers, though the authors caution that agent errors may be harder to spot. A 2025 trial by METR found that 16 experienced developers took 19% longer with AI tools after expecting to save 24%, suggesting that perception of review time is itself a risk.
The broader lesson is that productivity gains from agents depend on designing systems that help people judge output efficiently. Without such systems, the reading tax will consume the gains, and human oversight will become a formality rather than a safeguard.
3
