🔍 Read the full analysis: The Decision-First Way I Work With AI Agents 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 describes a four-rule approach to using AI agents: account for the work their output creates, test ideas without committing to every one, set stop conditions for recurring tasks, and limit how often a coordinating agent seeks attention. The approach is informed by Anthropic’s study of its own workforce, which the source says found substantial agent wait time and a growing burden of managing and reviewing output.
Thorsten Meyer published a four-rule approach to working with AI agents on October 27, arguing that people should manage agents around the decisions and follow-up work they create, rather than the speed at which they produce output. His framework draws on an Anthropic study of its own workforce, which the source reports found that agents often waited for human input and that reviewing their output could consume substantial staff time.
Meyer’s rules are to count the human work that follows an agent’s output, use agents for small tests without turning every test into a permanent project, set an end point before making a task recurring, and give a coordinating agent a defined attention budget, an approach that also raises questions about who answers for AI agents. He proposes that such an agent remain quiet by default and raise no more than three decisions a day, each with a recommendation, supporting evidence and the consequence of taking no action.
The article reports that Anthropic’s study found that, by August 2026, employee-directed agents logged six times as many hours as the employees directing them. It says agents spent about 80 percent of session time waiting for human input. The study’s estimate, as relayed by Meyer, was that reading all agent output would take non-technical employees about half their working time.
Meyer also cites a rise in the share of sessions devoted to “meta-work,” mainly agent management, from 10 percent in fall 2025 to 22 percent in spring 2026, underscoring the importance of setting limits for agents. Written standing instructions for agents reportedly grew from a few hundred to 8,500 over a year. The figures are presented as findings from one company, not as a measure of the wider economy.
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
Meyer’s central point is that fast, inexpensive generation does not remove the work needed to read, verify and act on what an agent produces. If people cannot use an output, or if checking it takes longer than the task it was meant to help with, adding agents may increase workload rather than reduce it.
The framework makes the human decision process part of the design. Before assigning work, Meyer suggests identifying who will review it, who will act on it and whether the resulting work is wanted. That distinction matters for businesses adopting agents: a useful proposal can lead to paid work, while a polished but incorrect estimate can create risk and additional checking.
The proposed attention budget also addresses notification and review fatigue. Restricting an agent’s interruptions to a small number of actionable decisions could help people retain oversight without treating every status update as urgent. These are Meyer’s recommendations, not reported results from a controlled test of the four rules.
Products Worth Considering
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Dell UltraSharp 27 Thunderbolt Hub Monitor (U2726DER)
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What Anthropic’s Figures Measure
The data Meyer cites came from Anthropic’s study of its own employees, published October 16, 2026, according to the article. The source cautions that staff adopted the tools by choice and that Anthropic’s working practices may not apply across the economy. The figures therefore describe the company’s observed use, rather than establishing a general rate for all businesses or agent systems.
The article also includes examples from Anthropic customer accounts. Dan Ninerell, who runs Modern Classical Chefs in South Jersey, said Claude helped him draft catering proposals and attributed $20,000 in revenue over six weeks to that work. This is Ninerell’s account, published by Anthropic, and not an independently verified outcome or a guarantee of results for other users.
Meyer describes contractor Tony Severino finding that Claude had used a floor-area multiplier rather than actual wall measurements when reading blueprints for a bid. Severino now asks it to “show me your work.” Meyer also cites business owner Rick Smith asking Claude to distinguish assumptions from known information. The examples support the article’s recommendation to check consequential calculations and make uncertainty visible.
““an army of Claudes, with every one asking for my attention.””
— An engineer quoted in Meyer’s article
Limits of the Workforce Evidence
The article does not provide the study’s full methodology, sample size or detailed definitions for measures such as agent “hours” and session waiting time. Readers therefore cannot assess those figures fully from Meyer’s account alone. Anthropic’s own workforce is also a single-company sample, and the source says employees adopted the tools by choice.
It is also unclear whether Meyer’s four rules have been tested as a package or how much time they save in different types of work. The customer revenue figure is attributed to Ninerell and published by Anthropic; the article does not independently verify it. The source provides recommendations and examples, not evidence that every business will get the same results.
How to Apply the Four Rules
Meyer’s suggested next step is to use three questions before assigning an agent: Who will read the result? Who will act on it? Do I want the work it creates? For recurring jobs, he recommends setting a finish date, stopping when there is nothing new to report and pausing a task if nobody reads its output.
His coordinating-agent rule offers another practical test: keep routine updates silent and surface only a limited number of decisions, each accompanied by a recommendation, evidence and the cost of inaction. The article does not announce a new product or a formal rollout. It presents an operating framework that readers can try with existing agents and adjust based on whether the resulting work is used.
Key Questions
What is Meyer’s decision-first approach to AI agents?
It is a four-rule framework: count the human work after an agent responds, run small tests without committing to every idea, set stop conditions for recurring tasks, and limit how often a coordinating agent demands attention.
What did the Anthropic study report?
As summarized in Meyer’s article, employee-directed agents logged six times the hours of the employees directing them by August 2026, spent about 80 percent of session time waiting for human input, and generated enough output that reading it could take non-technical staff about half their work week. These are findings from Anthropic’s own workforce, not an industry-wide estimate.
Why does Meyer say more agents can create more work?
Agents can make output quickly, but people still need to read, check and act on it. If the output is not useful or creates more review work than it saves, using more agents can add to a person’s workload.
What should a coordinating agent send to its user?
Meyer recommends silence by default and no more than three decisions a day. Each decision should include a recommendation, the supporting evidence and what may happen if the user does nothing.
Are the reported business results guaranteed?
No. The article recounts customer claims published by Anthropic, including Ninerell’s reported revenue from catering proposals. Those accounts are not guarantees or independently verified predictions of results for other users.
Source: ThorstenMeyerAI.com
