🔍 Read the full analysis: Tired Of Babysitting AI Agents? Ask For Decisions on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
An October 16 Anthropic study of its own workforce found employee-directed AI agents logged six times the hours of their human directors by August 2026, while spending about 80% of session time waiting for people. A commentary published October 27 uses those findings to argue that agent users should budget for human review, limit recurring work and ask coordinating agents to surface only a few decisions.
A commentary published October 27 argues that people using AI agents should manage the attention those systems demand, after an Anthropic study of its own workforce found that employee-directed agents spent about 80% of session time waiting for human input. By August 2026, the agents in those sessions had logged six times as many hours as the employees directing them, according to the study as summarized by Thorsten Meyer.
The figures describe Anthropic employees’ work with agents, not a broad survey of workplaces. The study was published October 16. Meyer cautions that Anthropic’s staff chose to adopt the tools and that the company’s work may not reflect other employers. The reported ratio compares logged agent hours with the hours of the employees directing them; it does not mean agents completed six times as much useful work.
The source says agents spent about 80% of session time waiting for human input. It also reports an estimate that reading all agent output would take non-technical employees about half their working time. The share of sessions devoted to “meta-work,” described mainly as managing agents, rose from 10% in fall 2025 to 22% in spring 2026. Written standing instructions grew from a few hundred to 8,500 over a year, according to the study’s figures cited by Meyer.
Meyer’s practical recommendations are to count the work required to review and use an agent’s output; test ideas without committing to every promising result; set an end date and stopping rules for recurring tasks; and give a coordinating agent an attention budget. He suggests a default of silence and no more than three decisions a day, each presented with a recommendation, supporting evidence and the consequence of taking no action.
The new bottleneck
Agents don't get tired. Your attention does.
Anthropic measured how its own staff work with AI agents. The agents now do far more hours. The limit is the person who has to read, check and decide.
An agent drafts the report, the code or the summary in minutes.
- Reading
- Checking
- Deciding
- Fixing
- Explaining it to others
Illustration of the idea, not a measurement.
“An army of Claudes, with every one asking for my attention.”An Anthropic engineer, quoted in the study
Source: Anthropic Economics team, How employment is changing at Anthropic, 16 October 2026. Employee-directed interactive sessions; human time counted as 40 hours a week per head. One company studying its own staff, who chose to adopt the tools.
Powered by Thorsten Meyer AI
Human Review Sets the Real Cost
The findings highlight a gap between the apparent cost of generating information and the cost of acting on it. An agent may produce a report quickly, but a person still has to read, verify and decide what to do. If many agents generate updates or proposals at once, their combined output can compete with the work they were meant to save time on.
That distinction matters to companies adopting agent tools: the number of tasks completed or hours agents run may not show whether people are saving time. The relevant measure may include human review and follow-up. Meyer’s rules are recommendations, not tested results from a controlled evaluation; readers should treat them as a framework for managing workloads rather than proof that a particular setup will improve productivity.
The commentary also distinguishes useful output from merely polished output. Meyer cites a contractor who found that an agent had used a floor-area multiplier instead of actual wall measurements when preparing a bid. Asking the system to show its work, he says, can focus human checking on the parts where an error could matter. The example is an anecdote published in Anthropic’s customer material, not an independent assessment of accuracy.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
DJI Osmo Mobile 8 3-Axis Smartphone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
What Anthropic Measured
Anthropic’s study examined its employees’ use of AI agents over time. The figures quoted in Meyer’s October 27 article include agent activity through August 2026 and changes in management-related work from fall 2025 to spring 2026. The study’s setting matters: employees opted into using the tools, and the company’s work patterns may not match those of other organizations.
Meyer also draws on a familiar feature of agent-based work: delegating a task does not remove human responsibility for its result. He recommends asking, before starting, who will read the output, who will act on it and whether the resulting work is worth taking on. For recurring tasks, he advises setting an end date, stopping when there is nothing new to report and pausing a process if nobody is reading its output.
The article includes customer examples from Anthropic’s own material. A chef says Claude helped draft proposals linked to revenue; a contractor describes checking the system’s measurements on a bid. These are vendor-published customer accounts. They illustrate possible uses but do not establish typical financial returns or error rates.
“human attention is now the bottleneck for agent output.”
— Thorsten Meyer, summarizing the study’s conclusion
How Far the Findings Apply
The figures cited here come from one company’s workforce. The source does not establish whether the same agent-to-employee hours, waiting time or review burden apply across industries, job types or other AI systems. Anthropic’s employees adopted the tools by choice, and Meyer notes that some of the company’s working practices may not be transferable.
The cited figures also do not, on their own, show whether agent use raised or lowered overall productivity, improved work quality or reduced total staff time. The article does not provide a comparison with a group that did not use agents. The business outcomes described by customers are individual accounts, and the source does not independently verify them. It is also unclear how the study defined or measured an agent’s “hours” and estimated time spent reading output beyond the summary provided.
Track Review Before Scaling
Meyer’s next step for agent users is to check the work created after generation, not just the speed of the initial task. Before launching an agent, teams can identify who will review the result and what action is expected. For recurring tasks, they can set a review date and a clear condition for stopping or pausing the work. Coordinating agents can be instructed to raise only decisions that need attention, with evidence and a recommended action.
The source does not identify a follow-up release date or a new Anthropic study. Whether these practices reduce workload remains a question for teams to test against their own tasks, including review time, errors, useful decisions and abandoned outputs. Broader evidence would be needed to determine how widely the findings apply.
Key Questions
What did Anthropic’s study find about AI agents?
As summarized by Thorsten Meyer, by August 2026 employee-directed agents had logged six times the hours of the employees directing them and spent about 80% of session time waiting for human input. The findings concern Anthropic’s own workforce.
Does the study show agents made employees more productive?
Not from the figures described in the source. Agent hours and waiting time do not by themselves establish changes in productivity, work quality or total time saved. The article does not provide a non-agent comparison group.
What does Meyer recommend to reduce agent-related workload?
He recommends counting review and follow-up work, treating early experiments as tests rather than automatic commitments, setting end dates and stop rules for recurring tasks, and limiting a coordinating agent to a small number of decisions that need human attention.
Are the customer revenue examples independently verified?
No independent verification is described. The chef’s revenue figure is his own account, published in Anthropic’s customer material. It should not be read as a typical or guaranteed result.
Source: ThorstenMeyerAI.com
