🔍 Read the full analysis: Less Checking, More Deciding: My Four Rules For AI Agents on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer argues that AI agents can shift work from producing output to reading, checking and acting on it. Drawing on Anthropic’s study of its employees, he recommends counting the human work an agent creates, treating experiments as experiments, setting limits on recurring tasks and keeping agent alerts quiet by default.
Thorsten Meyer has published four rules for managing AI agents, arguing that the main constraint is often not how quickly agents produce work but how much human attention their output requires. His guidance follows an Anthropic study of its own workforce, which reported that employee-directed agents spent about 80 percent of session time waiting for people and generated more output than staff could easily review.
Meyer’s first rule is to count the work that follows an agent’s output: who will read it, who will check it, and who will act on it. He argues that fast, inexpensive generation can still create a costly task if several people must review the result. The key question, in his account, is whether the work created by a successful output is work someone actually wants to do.
His other rules are to run many small experiments but keep only the ones that prove useful; decide when a recurring agent task will end before making it repeat; and give a coordinating agent a limited attention budget. Meyer proposes silence as the default, with no more than three daily decisions raised to a person, each accompanied by a recommendation, supporting evidence and the consequence of doing nothing.
The article also emphasizes checking work where errors matter. Meyer cites a contractor who found that Claude had used a floor-area multiplier rather than actual wall measurements when reading plans for a bid. The contractor subsequently asked the system to “show me your work.” The example illustrates Meyer’s warning that a polished answer may still contain a consequential mistake.
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 argument is that adding agents does not automatically reduce workload. It can move effort from drafting or research to reading, verification and follow-up. That matters to teams adopting agents because the time saved by generating a report or proposal is not the same as time saved overall if people must still assess every result.
Anthropic’s figures offer a reason to track that human work rather than counting only agent activity. The study estimated that reading all agents’ output would take non-technical employees about half their working time. That is an estimate reported in the study, not a measure of every worker’s experience. Meyer’s rules respond by limiting what agents produce for review and by requiring a clear purpose for repeat tasks.
The approach also has implications for accountability. Asking an agent to expose its reasoning or identify assumptions can help a person focus checks on risky details. It does not remove the need for review, but it may make review more targeted.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What Anthropic Measured
Anthropic’s workforce study, published on October 16, 2026, examined how its own employees used agents. As summarized by Meyer, by August 2026 employee-directed agents were logging six times as many hours as the employees directing them. The study also found that the share of sessions spent on “meta-work,” described mainly as managing agents, rose from 10 percent in fall 2025 to 22 percent in spring 2026.
Those findings have limits. The research concerned one company, whose employees adopted the tools by choice, and the study’s authors cautioned that Anthropic’s working practices may not apply across the economy. The numbers should be read as evidence about that workforce, rather than a universal forecast for AI use.
Meyer writes from his own experience running agents for publishing and software work. His four rules are recommendations, not results from a controlled trial. The article also refers to customer accounts published by Anthropic, including a chef who credited Claude-assisted proposals with helping him earn $20,000 over a month and a half. That is the business owner’s account in vendor material, not an independently verified finding.
“an army of Claudes, with every one asking for my attention.”
— An engineer quoted in Meyer’s article
How Far the Findings Apply
The Anthropic study describes one company’s workforce, and its results may not represent organizations with different roles, workflows or levels of agent adoption. The source material does not establish that the four rules reduce total work across a range of businesses, nor does it provide a measured comparison of outcomes for users who apply them.
Several figures and business examples come through Meyer’s account of Anthropic research or customer stories. The customer revenue figure is the owner’s claim as published by Anthropic, rather than an independently audited result. It is also unclear how the suggested three-decision daily limit would work across teams with different demands or time-sensitive responsibilities.
Testing the Rules in Daily Work
Meyer presents the rules as practices people can apply with agents they already use, rather than as a new product or formal standard. His proposed next step for readers is to review each agent task: identify who will handle the output, keep experiments small, assign recurring tasks a stopping point and limit notifications from coordinating agents.
Whether those practices reduce the attention burden will depend on the task and the organization. The source material does not announce a follow-up study or a planned measurement of the rules’ results. For now, Anthropic’s workforce findings and Meyer’s recommendations provide a case for tracking the human time spent reviewing and managing agents alongside the work those systems produce.
Key Questions
What are Thorsten Meyer’s four rules for AI agents?
Meyer recommends counting the human work an agent creates, running small tests without turning every test into a permanent project, setting an end point for recurring tasks, and keeping a coordinating agent quiet unless it has a limited number of meaningful decisions to raise.
What did Anthropic’s study report about agents waiting?
As summarized by Meyer, agents in employee-directed sessions spent about 80 percent of session time waiting for human input. The study examined Anthropic’s own workforce, so that figure should not be treated as a universal rate.
Does the study prove that AI agents increase workload?
No. The study describes agent use at Anthropic and reports measures such as time spent waiting and managing agents. Meyer argues that reviewing output can create additional work, but the material does not establish that every organization experiences a net workload increase.
Why does Meyer recommend asking an agent to show its work?
He cites a contractor who found that Claude had used an incorrect measurement approach when reading plans. Asking for the work or assumptions can help a person check important details, though it does not make the result automatically correct.
Source: ThorstenMeyerAI.com
