🔍 Read the full analysis: A Four-Rule Framework For Getting Decisions From 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
An October 27 article by Thorsten Meyer proposes four rules for managing AI agents: account for the work their output creates, test broadly but commit selectively, set stopping conditions for recurring jobs, and limit how many decisions a coordinating agent sends to a person. The framework draws on Anthropic’s study of its own employees, which reported extensive agent waiting time and a growing share of work spent managing agents; those findings come from one company and are not a universal measure.
Thorsten Meyer published a four-rule framework for managing AI agents on October 27, arguing that people need to limit the review and follow-up work agents create, not just make agents faster. The article cites an Anthropic study of its own workforce, which found that employee-directed agents spent about 80 percent of session time waiting for human input.
Meyer’s rules are to count the human work that follows an agent’s output, run multiple small tests but commit to only a few, set an end date and stopping conditions before making a task recurring, and give a coordinating agent an explicit attention budget. Under the proposed default, that agent stays silent unless it has a decision to raise, and brings no more than three decisions a day, each with a recommendation, evidence and the consequence of taking no action.
The article reports that Anthropic’s study found employee-directed agents logged six times as many hours as the employees directing them by August 2026. It also estimates that reading all agent output would take non-technical employees about half their working time. These are findings attributed to the study, not measurements of all workplaces.
The framework focuses on the costs after an agent produces an answer: someone still has to read it, check it and decide what to do. Meyer’s first rule asks users to identify those people and tasks before starting an agent. He also recommends asking agents to show their work or distinguish assumptions from known information, particularly when a polished answer could conceal a 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
Why Agent Output Needs Limits
The framework addresses a mismatch between the low cost of generating AI output and the continuing cost of human attention. An agent can produce a report quickly, but its speed does not remove the need to review the report, verify its claims or act on its recommendations. If several agents each send results that appear useful, their combined demands may become a substantial part of someone’s workload.
That matters to organisations adopting agents for research, planning and recurring administrative work. The proposed rules treat review capacity as a resource to manage: teams should ask whether an output has a reader and an intended use, and whether a recurring task still deserves to run. Limiting alerts and decisions could also help people focus on consequential issues rather than routine status updates.
These are practical recommendations, not proof that the framework will improve productivity. Meyer presents them as rules he follows in publishing and software work. Readers and employers would need to test whether the approach reduces unnecessary review without causing important decisions or risks to be missed.
Products Worth Considering
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Jackery Solar Generator 1000 v2 with 200W Solar Panel
Bose S1 Pro+ Portable PA Speaker
As an affiliate, we earn on qualifying purchases.
What Anthropic Measured
The article cites an Anthropic study of the company’s own workforce, published on October 16, 2026. According to Meyer’s account of the findings, agents spent about 80 percent of session time waiting for a human response. 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. The article also says written standing instructions for agents grew from a few hundred to 8,500 over a year.
Meyer cautions that the study is a single-company signal, not a law that applies across the economy. The study’s authors, as described in the article, noted that Anthropic employees adopted the tools by choice and that aspects of the company’s work may not be transferable. That caveat matters: the figures describe a specific workforce and setting, rather than a representative survey of businesses or workers.
The article illustrates its first rule with accounts from Anthropic customer stories. A private chef said AI-assisted proposals helped bring in revenue, while a painting contractor described catching an area-calculation error and then asking the agent to show its work. These are individual accounts published by the vendor, not independent evaluations of typical results.
“an army of Claudes, with every one asking for my attention.”
— An engineer quoted in Meyer’s article
How Widely the Findings Apply
The reported measurements come from Anthropic’s own employees. The article does not establish whether other companies would see the same agent usage, waiting time or management burden. It also does not provide enough detail here to assess how the study defined a session, calculated reading time or selected the employees included.
It remains unclear whether Meyer’s four rules have been tested against other approaches or whether they produce measurable gains in time, output quality or decision-making. The article offers recommendations and examples, but no controlled comparison. Nor does it quantify how often a strict attention limit might delay a time-sensitive warning or prevent an agent from raising a useful update.
Testing the Rules in Practice
Meyer’s proposed next step is to apply the rules to each agent before putting it to work: identify who will read and act on its output, decide which experiments merit further investment, set an end date or a “nothing new” stop for recurring tasks, and specify how many decisions the coordinating agent may raise. The framework also calls for pausing a recurring task if nobody is reading its results.
The article provides prompts and a three-question check for readers to use, but the source material does not report a formal trial or a scheduled follow-up. Whether the approach helps will depend on how teams weigh fewer interruptions against the risk of missing useful information.
Key Questions
What are the four rules for managing AI agents?
Count the human work an agent’s output creates; run small tests but commit selectively; define an end date and stopping conditions for recurring tasks; and set an attention budget for any agent that coordinates other agents.
What did Anthropic’s workforce study report?
As described by Meyer, employee-directed agents logged six times the hours of the employees directing them by August 2026, and spent about 80 percent of session time waiting for human input. The findings concern Anthropic’s workforce, not workplaces generally.
What does an attention budget mean here?
Meyer suggests that a coordinating agent should default to silence and bring a person no more than three decisions a day. Each decision should include a recommendation, supporting evidence and what may happen if no action is taken.
Are the framework’s results independently verified?
The source presents the rules as Meyer’s recommendations and examples, not as the results of a controlled test. The cited workforce data come from Anthropic’s study of its own employees, and the customer accounts are published by Anthropic.
Source: ThorstenMeyerAI.com
