🔍 Read the full analysis: Stop Checking In: Get AI Agents To Return Decisions on ThorstenMeyerAI.com
TL;DR
An October 2026 article draws on Anthropic’s study of its own workforce to argue that AI agents can create more human review and coordination work than expected. It recommends accounting for follow-up work, keeping only useful experiments, setting stop conditions for recurring tasks and limiting agent alerts.
A ThorstenMeyerAI.com article published October 27 argues that people using AI agents should limit the work those agents send back for review, drawing on an Anthropic study of its own employees that found agents spent about 80% of session time waiting for human input. The article proposes four practices: count the review and action work an agent creates, test ideas without adopting every result, set stop conditions for repeated jobs, and limit how often a coordinating agent asks for attention.
The figures in the article come from Anthropic’s study of its own workforce, published October 16, 2026. By August, agents directed by employees were logging six times as many hours as the employees directing them. The article says about 80% of session time was spent waiting for a person to respond. Those figures describe Anthropic’s measured sessions, not all AI agent use.
The article also reports the study’s estimate that reading all agent output would take non-technical employees about half their working time. It says the share of sessions devoted to “meta-work,” mainly managing agents, rose from 10% in fall 2025 to 22% in spring 2026. Written standing instructions grew from a few hundred to 8,500 in a year, according to the source article’s account of the study.
Thorsten Meyer’s four proposed rules are practical guidance, not findings tested by Anthropic. They are to identify who will read, check and use an output; use agents for low-cost exploration but adopt only results that prove useful; give recurring tasks an end date and stop them when there is nothing new or no one is reading the output; and set an attention limit for a coordinating agent. The article suggests a default of silence and no more than three daily decision requests, each with a recommendation, evidence and the consequence of taking no action.
Anthropic study
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.
ReadingCheckingDecidingFixingExplaining 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
When Agent Output Becomes Human Work
The central issue is a mismatch between the cost of producing information and the cost of using it. An agent may generate a report quickly, but a person still has to read it, check it and decide what to do. If agents make it easier to produce more output than people can review, adding agents may shift work rather than remove it.
That matters to teams adopting agents for research, proposals, monitoring or software tasks. A stream of individually reasonable updates can create a steady demand for attention, while unchecked outputs can be misleading. The article’s suggested limits aim to preserve the useful parts of delegation while making human review and follow-up visible costs, rather than treating generated material as free.
The Anthropic Workforce Measurements
The study cited by the article covers Anthropic employees using the company’s agents. The source article cautions that the findings are a signal, not a general rule: employees adopted the tools by choice, and the company’s work may not match how other organizations operate. The measured time and estimates should be read within that scope.
The article also points to examples Anthropic published about business customers. Chef Dan Ninerell said Claude helped him prepare proposals and attributed $20,000 in revenue over a month and a half to that work; this is his account, published by the vendor, not an independently verified result. Contractor Tony Severino described catching a floor-area calculation problem in an AI-assisted bid and subsequently asking the system to show its work. These examples illustrate the article’s distinction between output that supports wanted work and output that requires costly checking.
“I made $20,000 in the last month and a half using Claude to do these professional proposals”
— Dan Ninerell, owner of Modern Classical Chefs, as quoted in Anthropic customer material
How Far the Findings Apply
The figures do not establish how much time agents consume across other companies or industries. Anthropic’s employee sample was self-selected, according to the article, and the study authors cautioned that the company’s work may not be representative. The source material does not provide enough detail here to assess the study’s full methods or how its estimates were calculated.
It is also unclear whether the article’s four rules have been tested in a controlled setting or how much they reduce review time in practice. Its customer examples are attributed accounts, not evidence that other users should expect comparable results. The suggested limit of three decision requests a day is Meyer’s recommendation, not an Anthropic standard.
Testing Limits in Daily Work
The article offers prompts and a three-question check for readers to use with agents, but the source material does not identify a scheduled follow-up study or an independently measured test of those practices. For now, its advice is to set boundaries before deploying an agent repeatedly: specify who owns the output, when the task ends and what should happen when there is nothing new.
For organizations, the next useful measure is not only how quickly agents produce work, but how much time people spend reviewing, correcting and acting on it. Whether that balance improves will depend on the task and the agent’s reliability; the Anthropic figures alone do not settle that question.
Key Questions
What did the Anthropic study find about agent waiting time?
As reported in the article, agents in employee-directed sessions spent about 80% of session time waiting for human input. The figure comes from Anthropic’s own workforce study and should not be treated as a measure of all agent use.
Did agents log more hours than Anthropic employees?
The article says that by August 2026, employee-directed agents logged six times as many hours as the employees directing them. This is a measurement reported for Anthropic’s workforce, not a general industry statistic.
What four practices does the article recommend?
It recommends counting the human work after an agent responds, testing ideas without turning every test into a permanent project, setting end and stop conditions for recurring tasks, and limiting how often a coordinating agent interrupts people.
Are the article’s rules proven to reduce workload?
The source presents them as Meyer’s practical recommendations. It does not provide controlled evidence that the four rules reduce workload across organizations.
Source: ThorstenMeyerAI.com