🔍 Read the full analysis: AI Agents And Decision-Making: Four Rules That Guide My Workflow 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
Thorsten Meyer’s October 27 essay argues that AI agents can shift work rather than remove it, because people still need to review their output and make decisions. He recommends counting follow-up work, treating agent projects as tests before expanding them, setting end conditions for recurring tasks and limiting the decisions a coordinating agent sends to a person.
Thorsten Meyer published four rules on October 27 for managing AI agents, arguing that human review and follow-up can turn supposedly low-cost automation into more work. His essay draws on an Anthropic study of its own workforce and recommends budgeting for the attention agents require, testing ideas without automatically expanding them, setting stop conditions for recurring tasks and limiting the decisions a coordinating agent sends to a person.
Meyer says that before starting an agent, users should ask who will read the result, who will act on it and whether they want the work that follows. His argument is that generating a report quickly does not make the report cheap if several people must spend hours reviewing it. He advises checking consequential outputs, including asking an agent to show its reasoning or distinguish assumptions from known information.
His second rule is to let agents test multiple ideas while committing to only those that demonstrate value. The third is to define an end date, a “nothing new” stop condition or a pause for any recurring task. The fourth is to give a coordinating agent an attention budget: default to silence and surface no more than three daily decisions, each with a recommendation, supporting evidence and the consequence of taking no action.
The figures Meyer cites come from Anthropic’s study of its employees, published October 16. The essay reports that, by August 2026, employee-directed agents logged six times as many hours as the employees directing them, while agents spent about 80% of session time waiting for human input. Meyer also cites the study’s estimate that reading all agent output could take non-technical employees about half their work week. These are findings from one company’s workforce, not a general measure of agent use.
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
Meyer’s core point is that agent output still creates human work. People may need to read, verify, prioritize and act on what systems produce; those tasks can outweigh the time saved during generation. A polished result may also be misleading if its calculations or assumptions are wrong, making review especially important where errors affect bids, spending or client commitments.
The proposed rules offer a practical way to judge whether automation is helping: count the work that follows, and stop or limit tasks whose output nobody uses. That approach matters to teams adopting agents because the number of agents or tasks completed does not, by itself, show that people have saved time. Meyer’s framework treats attention as a finite resource to budget alongside money and computing capacity.
Popular right now
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Breville Barista Pro Espresso Machine with Grinder, Brushed Stainless Steel
Roborock Robot Vacuum and Mop with LiDAR Navigation and 10,000 Pa Suction
As an affiliate, we earn on qualifying purchases.
Anthropic’s Workforce Findings
Meyer’s essay applies the Anthropic workforce study to his publishing and software work. He describes a familiar pattern: an agent sends updates, a summary still needs to be read, and a completed task produces suggestions that require decisions. In his account, each message can seem reasonable alone, while the combined stream becomes a separate management job.
The study’s reported indicators include agent sessions spending about 80% of their time waiting for a person and the share of sessions devoted to “meta-work,” mostly managing agents, rising from 10% in fall 2025 to 22% in spring 2026. Meyer also reports that written standing instructions for agents grew from a few hundred to 8,500 over a year. The study’s authors cautioned, as described in the essay, that Anthropic staff adopted the tools by choice and that the company’s working practices may not generalize to other workplaces.
Meyer also cites customer accounts published by Anthropic. A private chef said Claude helped draft proposals that he associated with $20,000 in business over a month and a half; that is the owner’s account, not an independently verified outcome. A painting contractor described catching an agent’s incorrect use of a floor-area multiplier and said he now asks it to “show me your work.” These examples illustrate Meyer’s recommendations but do not establish typical results.
“an army of Claudes, with every one asking for my attention.”
— An engineer quoted in Meyer’s essay
Limits of the Available Evidence
The Anthropic figures describe one company’s workforce, and the source material says the study’s authors warned against treating its findings as representative of the wider economy. The reported ratios and time estimates may not apply to workplaces with different tasks, agent tools or review practices. The essay does not provide enough detail to assess the study’s methods independently.
Meyer presents the four rules as guidance from his own workflow, not as results from a controlled test. The source material does not establish how much time or money the rules save, whether the suggested three-decision limit works for every team, or how often agent output leads to errors. The customer examples are attributed accounts, and their outcomes should not be read as typical or independently confirmed.
How Teams Can Test the Rules
Meyer’s recommendations point to a local test rather than a single measure of agent activity: teams can track how long people spend reading, checking and acting on outputs, and whether those outputs lead to useful decisions. For recurring work, they can set a review date or a stop condition in advance, then pause tasks whose results are not being used.
The essay does not announce a product launch or a formal workplace policy. Its next step is for readers to apply the rules to their own workflows and judge whether agents reduce total effort, rather than only the time required to generate a result. Whether the approach works beyond Meyer’s examples remains an open question.
Key Questions
What are Meyer’s four rules for using AI agents?
He advises users to count the human work that follows an agent’s output, run small tests without turning every test into an ongoing project, set end conditions for recurring tasks and give coordinating agents a limit on how much attention they can demand.
What did the Anthropic study report?
As summarized by Meyer, employee-directed agents logged six times the hours of the employees directing them by August 2026 and spent about 80% of session time waiting for human input. The figures concern Anthropic’s own workforce.
Does the study show that agents increase workload at every company?
No. The source material says Anthropic’s authors cautioned that its employees chose to adopt the tools and that the company’s practices may not generalize. The findings are a signal about one workforce, not proof of a universal effect.
What does an agent attention budget mean?
Meyer suggests that a coordinating agent default to silence and send a person at most three decisions a day. Each should include a recommendation, evidence and what may happen if no action is taken; he presents this as his workflow rule, not a tested universal standard.
Source: ThorstenMeyerAI.com
