🔍 Read the full analysis: Are You Overseeing Your AI Agents Too Closely? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Anthropic’s October 16, 2026 study of its own workforce found employee-directed agents logged six times as many hours as their human directors by August, while spending about 80% of session time waiting for human input. The findings suggest that reviewing and coordinating agent work can become a significant workload, though Anthropic cautions that its results may not represent other workplaces.
Anthropic’s study of its own workforce found that, by August 2026, employee-directed AI agents logged six times as many hours as the employees directing them, while spending about 80% of session time waiting for human input. The findings highlight a potential cost of agent use: people may spend substantial time answering agents, checking their work and deciding what to do with the output.
The study, published on October 16, examined how Anthropic employees worked with agents. The company estimated that even non-technical employees would need about half their working time to read all the output their agents produced. It also reported that the proportion of sessions devoted to meta-work, mostly managing agents, rose from 10% in fall 2025 to 22% in spring 2026.
Anthropic said written standing instructions for agents grew from a few hundred to 8,500 in a year. Taken together, the figures point to growing coordination and review demands at the company, rather than a simple measure of how much useful work agents completed. The study’s authors described human attention as a bottleneck for agent output.
The results are from one company’s workforce, not a representative survey of all workplaces. Anthropic said its employees adopted the tools by choice and that aspects of its work may not transfer to other industries. The numbers therefore provide a case study, not a universal estimate of how much time AI agents require.
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 Work Behind Agent Output
The findings matter because producing a result is only one part of an automated task. Someone may still need to read, verify and act on what an agent returns. When generating reports or proposals becomes cheaper and faster, the time needed to assess them may remain much the same. A stream of low-cost outputs can therefore create a real workload if no one decides which results deserve attention.
That distinction affects how organizations judge whether agents save time. Counting agent hours or completed tasks alone may miss the human effort spent supervising them. Managers may also need to account for the cost of correcting mistakes and maintaining recurring workflows. The Anthropic figures do not establish that agents increase workload overall, but they show why human review time belongs in the calculation.
For workers, the practical question is not only whether an agent can do a task, but whether its result creates work someone wants and has time to do. A draft that supports a paying client may be useful; a polished but inaccurate calculation could demand extra scrutiny. Setting limits on updates and asking agents to show evidence are among the approaches described in the source material, rather than findings tested by the study.
Products Worth Considering
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Anthropic’s Internal Agent Study
Anthropic’s measurements cover its own employees and their use of the company’s agents. The reported change in meta-work—from 10% in fall 2025 to 22% in spring 2026—describes the share of sessions spent mostly managing agents. It is not a measure of the share of all employee work devoted to AI, nor does it show whether that management produced better results.
The accompanying commentary by Thorsten Meyer applies the study to day-to-day agent management and proposes four practices: count the human work that follows an output, run experiments without turning every test into a permanent project, set stopping conditions for recurring tasks, and give a coordinating agent a limited attention budget. These are the author’s recommendations, not policies reported as adopted by Anthropic or conclusions proven by its workforce study.
The article also recounts customer examples published by Anthropic, including a chef who credited Claude-assisted proposals with helping his business earn money and a painting contractor who said he checks the system’s calculations. Those are vendor-published customer accounts, not independently verified results. They illustrate possible uses and checking practices but do not establish typical outcomes.
“Human attention is now the bottleneck for agent output.”
— Anthropic’s study authors
How Widely the Findings Apply
The study does not establish whether the same patterns occur in other companies, occupations or agent systems. Anthropic’s workforce is a single-company sample, and the company said employee adoption was voluntary and its working practices may not be readily transferable.
The reported agent hours and waiting time also do not, on their own, show how much useful work agents completed, how often their answers were wrong, or whether the time employees spent supervising them was outweighed by other savings. The estimate that reading output would take non-technical employees about half a work week is the study’s estimate; the available material does not specify enough detail to independently assess how it would apply to a particular job.
It remains unclear whether Anthropic will publish further measurements or whether companies outside Anthropic will track agent waiting, review time and resulting work in comparable ways. The customer examples in the accompanying article are individual accounts, not evidence that similar financial or productivity outcomes are typical.
Track Review Time and Outcomes
The immediate implication for employers is to measure more than how many tasks agents start or finish. Teams can track how much time people spend responding to agents, checking results and acting on them, alongside the usefulness and accuracy of those results. Such measurements would help distinguish work that agents genuinely reduce from work they shift to review and coordination.
Meyer’s recommendations offer one approach for individual users: decide who will use an output before commissioning it, stop recurring tasks when they no longer provide new information, and limit how often a coordinating agent asks for decisions. These remain practical suggestions in the commentary, not interventions whose effects were tested in Anthropic’s study.
Further evidence would be needed to learn whether the patterns hold across different workplaces and whether changes to agent design or management can reduce waiting and review demands. Until then, the study is a signal rather than a general benchmark; its reported measures should be read within the limits Anthropic described.
Key Questions
What did Anthropic’s study find about AI agents?
By August 2026, employee-directed agents at Anthropic logged six times as many hours as their human directors. The study also estimated that agents spent about 80% of session time waiting for human input.
Does the study show that AI agents make workers less productive?
No. The reported figures describe agent hours, waiting and management activity, but do not by themselves establish whether employees became more or less productive overall. Anthropic described the work as a study of its own workforce, not a universal result.
How much time might employees spend reading agent output?
Anthropic estimated that reading all agent output would take about half the working time of non-technical employees. That is the study’s estimate and should not be treated as a finding that applies equally to every workplace.
What does ‘meta-work’ mean in the study?
In the reported figures, meta-work refers mostly to managing agents. Its share of sessions rose from 10% in fall 2025 to 22% in spring 2026, according to Anthropic.
What can teams do to manage agent output?
The accompanying article recommends deciding who will use an output, setting stop conditions for recurring tasks and limiting unnecessary updates. These are the author’s suggestions, not practices shown by the study to work across organizations.
Source: ThorstenMeyerAI.com
