🔍 Read the full analysis: AI Agents Without Constant Check-Ins: Four Rules I Use on ThorstenMeyerAI.com
Get hardware and tech essentials 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
Thorsten Meyer’s October 27 guide recommends four ways to use AI agents without letting their output consume more human attention: count the work that follows, test ideas without committing to all of them, set stopping rules for recurring tasks, and limit what coordinating agents bring to a person. The article cites Anthropic’s study of its own workforce, which reported that agents in employee-directed sessions spent about 80% of the time waiting for human input.
Thorsten Meyer published a four-rule guide on October 27 for using AI agents without letting their updates and outputs become another job. The advice, based on his own publishing and software work, responds to an Anthropic study of its workforce that found agents in employee-directed sessions spent about 80% of their time waiting for a person.
Meyer’s central argument is that the cost of an agent is not just the time it takes to generate an answer. People still need to read, verify and act on the output. His four rules are to count that follow-up work before starting an agent; run inexpensive tests but commit only to ideas that prove useful; set an end date and stop conditions for recurring tasks; and give a coordinating agent a limited attention budget.
For coordinating agents, Meyer recommends silence as the default and a maximum of three decisions to bring to a person each day. Each item should include a recommendation, supporting evidence and what would happen if the person did nothing. For recurring work, he says to stop when there is nothing new and pause a task if nobody is reading its output.
Meyer cites Anthropic’s October 16 study for several measures of workload. By August 2026, employee-directed agents were logging six times as many hours as the employees directing them; the study estimated that agents spent around 80% of session time waiting for human input. It also estimated that reading all agent output would take non-technical employees about half their working time. These are findings from one company’s workforce, not a measurement of all workplaces.
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
The Human Cost of Agent Output
The guide focuses on a practical trade-off for people adopting AI agents: generating more work is not the same as completing more useful work. A quick report or proposal can still demand substantial review, checking and follow-up. If that labor is left out of the decision to automate, an agent may shift work onto a person rather than reduce it.
This matters to teams and small businesses that can now ask agents to produce material at low apparent cost. Meyer’s rules offer a way to count the human attention required before adding a new agent or recurring task. They also frame testing as separate from commitment: experimentation can be useful, but each successful trial need not become a permanent process to maintain.
The proposed limits are guidance, not findings tested by the cited study. Meyer describes them as rules he uses in his own work. Readers should treat the Anthropic figures as evidence about one organization, while the value of the rules in other settings will depend on the type of work, the quality of outputs and the time required to check them.
Products Worth Considering
LG 24G414B-B 24-Inch UltraGear Full HD IPS Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Anthropic’s Workforce Findings
Anthropic published its workforce study on October 16, 2026. As Meyer reports it, the study tracked the company’s use of employee-directed agents and found that their recorded hours had grown sharply relative to employees’ own hours. It also found that the share of sessions devoted to “meta-work,” mostly managing agents, rose from 10% in fall 2025 to 22% in spring 2026.
The study’s scope limits what its numbers can show. Meyer notes that Anthropic employees adopted the tools by choice and that parts of the company’s work may not carry over to other industries. The figures therefore provide a case study of agent use at Anthropic, rather than a universal estimate of how much time agents save or consume.
Meyer’s article also draws on Anthropic customer stories. One business owner said AI-assisted proposals helped generate revenue; another described checking an agent’s blueprint measurements after spotting an incorrect method. Those are individual accounts published by the vendor, not independently verified results. Meyer uses them to illustrate why review and evidence remain part of delegated work.
“an army of Claudes, with every one asking for my attention.”
— An engineer quoted in Meyer’s article
What the Data Cannot Establish
The study does not establish that AI agents will take up the same share of time in other companies, or that the same patterns apply across different jobs. The 80% waiting estimate describes sessions in Anthropic’s workforce study; it does not, on its own, measure whether the tools improved productivity or business outcomes.
It is also unclear from Meyer’s article how the four rules perform across a larger group of users, or how much time they save. The recommendations are presented as the author’s own working practices, not as results from a controlled trial. Customer revenue and productivity examples cited in the article are attributed accounts, and their measurement methods are not detailed there.
Testing the Rules in Practice
Meyer’s article presents the rules as practices readers can apply to their existing agents: identify who will read and act on an output, keep exploratory tests separate from ongoing commitments, write down when recurring work should stop, and cap the number of decisions a coordinator escalates. The approach calls for checking whether an agent’s output is actually used, rather than treating activity or message volume as success.
No formal follow-up study or wider test of the rules is announced in the source material. For now, the next step for readers is to compare the human time spent reviewing and managing agents with the work those systems help complete. Whether that balance improves will vary by task and will require evidence from each organization.
Key Questions
What are Thorsten Meyer’s four rules for using AI agents?
He recommends counting the human work that follows an agent’s output, running small tests without committing to every idea, setting end dates and stop conditions for recurring tasks, and limiting how many decisions a coordinating agent brings to a person.
What did Anthropic’s study report about agents waiting for people?
As reported by Meyer, the study estimated that agents in employee-directed sessions spent about 80% of session time waiting for human input. The finding concerns Anthropic’s own workforce, not all AI agent use.
Does the study show that AI agents save workers time?
Not by itself. The figures Meyer cites describe agent hours, waiting time, reading estimates and management work at Anthropic. They do not establish a general productivity effect across companies or occupations.
What should a coordinating agent send to a person?
Meyer suggests that it should default to silence and bring no more than three decisions a day, each with a recommendation, the evidence behind it and the consequence of taking no action.
Source: ThorstenMeyerAI.com
