🔍 Read the full analysis: How I Set Expectations For AI Agents In Four Rules on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer’s October 27 article proposes four rules for managing AI agents: count the human work their output creates, test widely but keep few projects, set stopping conditions for recurring tasks, and limit a coordinating agent’s demands on your attention. The guidance draws on an Anthropic study of its own workforce, which Meyer says should be treated as a signal rather than a universal measure.
Thorsten Meyer published four rules for managing AI agents on October 27, arguing that their output can create a substantial burden of reading, checking and follow-up work. The article draws on Anthropic’s study of its own employees, which Meyer says found that agents often waited for human input and generated more material than people could easily review.
Meyer’s rules are to count the human work that follows an agent’s output, run small experiments while committing to only a few, set an end date and stop conditions before making a task recurring, and give a coordinating agent a limited “attention budget.” Under his proposed default, that agent stays silent and raises no more than three decisions a day, each with a recommendation, evidence and the consequence of taking no action.
The article cites Anthropic’s October 16 research into its own workforce. According to the figures Meyer reports, by August 2026 employee-directed agents logged six times as many hours as the employees directing them. Agents spent about 80% of session time waiting for human input. Anthropic also estimated that reading all agent output would take non-technical employees about half their working time.
Meyer recommends asking three questions before launching a task: Who will read the result? Who will act on it? And is the work created by the result worth doing? He uses small-business examples from Anthropic’s customer stories to illustrate the distinction: a proposal can be useful when it leads to wanted paid work, but generated output still needs review. One contractor, he writes, found an early AI blueprint estimate used a floor-area multiplier rather than actual wall measurements.
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 Cost After an Agent Responds
The rules focus on a cost that can be hidden when generating text, analyses or proposals becomes fast: someone still has to read and verify the output. If agents make it easier to produce reports and recommendations, the number of items needing human attention may rise even when the time to create each item falls.
That matters to teams deciding whether to adopt agents or expand their use. Meyer’s approach treats attention as a limited resource, not an unlimited destination for automated work. Reviewing an agent’s reasoning is especially relevant when an answer looks polished but includes a mistaken calculation or assumption. The article’s practical message is to direct human checks toward the risks and actions that matter, rather than treating every generated result as ready to use.
The proposed limits also address recurring work. A routine that continues indefinitely can produce updates even when nothing has changed or nobody is acting on them. Meyer’s recommendation to set a stop condition and pause unread tasks offers a way to test whether an agent is still helping rather than simply adding notifications and upkeep.
Popular right now
De'Longhi Linea Classic Espresso Machine with Adjustable Milk Frother
Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151
Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew
As an affiliate, we earn on qualifying purchases.
Anthropic’s Workplace Findings
The research cited by Meyer examined Anthropic’s own workforce, rather than a representative sample of every industry or type of AI user. He reports that the share of sessions involving “meta-work,” mostly managing agents, rose from 10% in fall 2025 to 22% in spring 2026. Written standing instructions for agents grew from a few hundred to 8,500 over a year.
Meyer cautions that the study’s results should be read as a signal, not a general rule. The employees adopted the tools by choice, and the way Anthropic works may not translate to other workplaces. The article uses the findings to frame its four rules, but those figures do not establish that every organization will see the same agent workload or attention costs.
“human attention is now the bottleneck for agent output.”
— Anthropic research, as summarized by Meyer
Limits of the Workplace Data
The reported measures come from one company’s workforce, and the article does not establish whether the same ratios apply across other organizations, roles or agent systems. The study’s estimated reading time also describes the burden of reviewing all agent output; it does not show how much output employees actually read or how much proved useful.
The examples of business results are attributed accounts, not independently verified evaluations. The article does not provide comparative testing showing that Meyer’s four rules reduce workload, improve accuracy or produce better business outcomes. It presents them as his operating practices, supported by the problem described in the research.
Testing the Rules in Daily Work
Meyer’s suggested next step is to apply the rules before adding an agent or turning a one-off task into a recurring process. That means identifying who will review and use the result, deciding what evidence a human needs to check, and specifying when a recurring task should stop or pause.
For coordinating agents, he recommends a default of silence and a daily ceiling of three decision requests, each paired with evidence and a recommendation. Whether that limit works will depend on the person’s role and the risks involved; the article does not report a formal test of the threshold. The immediate test is whether the agent’s output leads to an action worth the attention it requires.
Key Questions
What are Meyer’s four rules for AI agents?
He says to count the human work after output, run small tests but commit to few, set an end date and stop conditions for recurring tasks, and limit how often a coordinating agent asks for attention.
What did Anthropic’s study report about agent use?
As summarized by Meyer, employee-directed agents logged six times the hours of their human directors by August 2026 and spent about 80% of session time waiting for human input. The research examined Anthropic’s own workforce.
Does the study show that all companies have the same problem?
No. Meyer says the findings are a signal, not a universal measure. The study focused on Anthropic employees, who adopted the tools by choice, and its working practices may not match those of other organizations.
Why does Meyer say cheap AI output can still be costly?
Generating a report or proposal quickly does not remove the time needed to read, check and act on it. Those follow-up tasks may become a significant workload when agents produce many results.
What does the article leave unproven?
It does not establish that the four rules reduce workload or improve outcomes across workplaces. The business examples are attributed accounts, and the article provides no controlled evaluation of the proposed practices.
Source: ThorstenMeyerAI.com
