🔍 Read the full analysis: My Four Rules For Delegating Work To AI Agents on ThorstenMeyerAI.com
Get the little things that make your day 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 article argues that AI agents can create more work for the people directing them when their output requires reading, checking and follow-up. Drawing on Anthropic’s study of its own workforce, he offers four rules for managing that burden, though the study’s findings may not apply broadly across workplaces.
Thorsten Meyer published four rules for delegating work to AI agents on October 27, arguing that the main constraint is often the human time needed to review and act on agent output. His advice follows Anthropic research on its own workforce, which reported that employee-directed agents logged six times as many hours as their human directors by August 2026 and spent about 80% of session time waiting for human input.
Meyer’s first rule is to count the work that comes after an agent produces an answer. Before assigning a task, he recommends identifying who will read the result, who will check it and who will act on it. An output that takes an agent minutes to produce may still require substantial staff time to verify or use.
His second rule is to let agents test multiple ideas without automatically turning every promising result into an ongoing project. The third is to set an end date and stopping conditions before making a task recurring, including stopping when there is nothing new to report or when nobody is reading the results.
The fourth rule is to give a coordinating agent an attention budget. Meyer says its default should be silence, with no more than three decisions brought to a person each day. Each item should include a recommendation, supporting evidence and the consequence of taking no action. These are the author’s proposed practices, not findings tested in the cited study.
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 rules address a gap between the low cost of producing AI-generated work and the human effort required to make that work useful. More reports, drafts and alerts do not automatically save time if people must read, verify and respond to them. Meyer’s central point is that attention is a limited resource, and a system that generates output without controlling its follow-up can shift rather than reduce workload.
That distinction matters for teams considering more agents or automated recurring tasks. A proposal, analysis or alert may be valuable when someone has a clear reason to use it. Without an owner or decision attached, the same output can become another item in an already crowded queue. Meyer’s suggested limits—small experiments, explicit stopping rules and fewer escalations—are intended to keep agent activity tied to work people actually want done.
Products Worth Considering
Withings Sleep Under-Mattress Sleep Tracking Pad
Therabody Theragun Mini 3rd Generation Massage Gun
Oral-B iO Series 5 Rechargeable Electric Toothbrush with Travel Case and 3 Brush Heads, Black
As an affiliate, we earn on qualifying purchases.
What Anthropic’s Study Measured
Anthropic’s workforce study, published October 16, examined how its employees used agents. The figures cited by Meyer include a rise in agent hours relative to employee hours: by August 2026, agents directed by employees logged six times the hours of those employees. The study also estimated that agents spent about 80% of session time waiting for a person, while reading all agent output would take non-technical employees about half their working time.
The study reported that 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 for agents increased from a few hundred to 8,500 over a year, according to the source material. These measurements describe Anthropic’s own employees and workplace, rather than a representative sample of all businesses.
Meyer also cites examples from Anthropic customer stories. Chef Dan Ninerell said Claude helped him produce catering proposals and attributed $20,000 in business over a month and a half to using the tool; that is his account, published by the vendor. Contractor Tony Severino described catching an early blueprint-reading error and now asking Claude to show its work. These anecdotes illustrate possible uses but do not establish typical results.
“Human attention is now the bottleneck for agent output.”
— Anthropic’s workforce study, as summarized by Meyer
How Far the Findings Apply
The Anthropic figures come from one company’s workforce. Meyer notes that employees adopted the tools by choice and that aspects of Anthropic’s work may not translate to other organizations. The study therefore offers a signal about possible attention costs, not a universal estimate of how agents affect productivity.
The source material does not provide enough detail to independently assess the study’s methods, sample sizes or how it defined and measured agent hours, waiting time and reading estimates. Nor does it establish that Meyer’s four rules reduce workload or improve results; they are recommendations based on his experience. The customer revenue example is also a vendor-published personal account, not an independently verified measure of the tool’s effect.
Testing the Rules in Practice
Meyer presents the rules as practices readers can apply to agents they already use. The next step for an individual or team is to track not only how quickly an agent completes a task, but also the time spent reviewing its output and the decisions it produces. For recurring work, teams can set a review date, a condition for stopping and an owner responsible for acting on results.
The article does not announce a formal trial or provide follow-up measurements showing whether these limits work. Any broader assessment would require comparing the full human workload—including checking and maintenance—before and after adopting agents, across more than one workplace.
Key Questions
What are Thorsten Meyer’s four rules for delegating to AI agents?
Count the human work that follows; run small tests but commit to few; set an end date and stopping conditions before making tasks recurring; and limit how many decisions a coordinating agent brings to a person.
What did Anthropic’s study report about agent waiting time?
As summarized by Meyer, agents in employee-directed sessions spent about 80% of session time waiting for human input. The finding comes from Anthropic’s own workforce and is not presented as a measure of every workplace.
Does the study prove that AI agents reduce productivity?
No. The figures cited point to time spent waiting, reading outputs and managing agents, but the source material does not establish a general productivity effect across organizations.
Why does Meyer recommend setting stopping rules?
He argues that recurring agent tasks can continue generating updates and review work even when they no longer provide useful information. An end date or a “nothing new” stop can limit that ongoing workload.
Is the $20,000 business figure independently verified?
The figure is chef Dan Ninerell’s own account, published in an Anthropic customer story. The material does not describe independent verification or show that the result is typical.
Source: ThorstenMeyerAI.com
