Stop Babysitting Your AI Agents: The 4 Rules I Use to Get Decisions, Not Updates

By Thorsten Meyer | October 27, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Anthropic recently measured what happened when its own staff started working with fleets of AI agents. By August 2026, the agents its employees directed were logging six times as many hours as the employees themselves. And in those working sessions, the agents spent around 80 percent of the time doing nothing at all. They were waiting for a person to answer.

One engineer put it in a single line: “an army of Claudes, with every one asking for my attention.”

That line describes a problem many of us now share. We hand work to agents to save time. Then the updates start. A watch reports that nothing happened. A finished task comes back with a list of new ideas. A summary agent sends a neat digest that still has to be read. Each message looks reasonable. Together they turn into a second job.

When I run agents for my publishing work and my software projects, I now follow four rules. They are simple, and they work with the agents you already have. Below are the rules, six prompts you can copy, and a three-question check for every agent you run.

Key takeaways

• The bottleneck is your attention, not the agent. Anthropic’s own study found that in employee-directed sessions, agents spent about 80 percent of the time waiting for a human. Reading all of their output would take even non-technical staff half their work week.
• Rule 1: Count the human work that comes after. Before you start an agent, ask who will read the result, who will check it and who will act on it. Cheap to make is not cheap to use.
• Rule 2: Run many small tests, but commit to few. Let agents explore ideas cheaply. Keep only what proves itself, and drop the rest without guilt.
• Rule 3: Decide when a job ends before you make it repeat. Every recurring task gets an end date, a “nothing new” stop and a pause when nobody reads it.
• Rule 4: Give your coordinating agent an attention budget. Its default is silence. It brings at most three decisions a day, each with a recommendation, the evidence and what happens if you do nothing.

Products Worth Considering

Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black

AmazonView Latest Price

Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black

AmazonView Latest Price

Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

Why more agents can mean more work

Before agents, effort did the filtering for us. A market report took a person a week, so you only asked for one when it mattered. Now an agent writes it in minutes. Asking feels free. Reading, checking and acting on it still costs what it always cost: your time.

Anthropic’s study of its own workforce, published on October 16, puts numbers on this. The figures come from one company studying its own staff, so treat them as a signal, not a law of nature. The authors say so themselves: their people adopted the tools by choice, and much of how Anthropic works cannot be copied across the economy. Still, the direction is clear:

• Six to one. By August 2026, employee-directed agents logged six times the hours of the people directing them.
• About 80 percent waiting. Most of a session’s clock time, the agent was idle, waiting for human input.
• Half a work week of reading. Even for non-technical employees, reading all of their agents’ output would take about half their working time, by the study’s estimate.
• Managing the managers. The share of sessions spent on “meta-work,” mostly managing agents, rose from 10 percent in fall 2025 to 22 percent in spring 2026. Written standing instructions for agents grew from a few hundred to 8,500 in a year.

The study’s own conclusion: human attention is now the bottleneck for agent output. That fits what psychologists have long known about us. We can hold only about four things in mind at once (Cowan, 2001). Agents do not get tired. We do.

Rule 1: Count the human work that comes after

The cost of an agent’s output is not the minutes it took to make. It is the time everyone spends on it afterward. A twenty-page proposal that three colleagues each need an afternoon to review is not a two-minute job. It is a claim on three afternoons that nobody budgeted.

So before I start an agent, I ask three questions. Who will read this? Who will act on it? And do I actually want to do the work it creates?

That last question matters most, and two small-business owners from Anthropic’s customer stories show why. Dan Ninerell runs Modern Classical Chefs, a private chef business in South Jersey. He uses Claude to draft catering proposals. “I made $20,000 in the last month and a half using Claude to do these professional proposals,” he told Anthropic. That figure is his own account, published by the vendor. The work that follows a good proposal is cooking for paying clients. That is work he wants.

Tony Severino, a Chicago-area painting contractor who has run his firm since 1984, uses Claude to read PDF blueprints floor by floor for his bids. On an early bid he caught Claude using a floor-area multiplier instead of real wall measurements. Now he asks Claude to “show me your work.” Another owner in the same Anthropic write-up, Rick Smith of Broadcast Blinds, asks Claude to flag when it is assuming and when it knows.

That is Rule 1 in practice. A polished answer with wrong math is worse than no answer, because it looks finished. Asking for the steps turns a long audit into a quick check. Your expertise goes where the risk is.

Two prompts for this:

Wrap this task up. Do not start anything new.
1. Send me the finished result in one message: what you made, where it is, and how you checked it.
2. List anything you could not finish and why. Do not do it now.
3. If you have ideas for more work, give me at most two, one line each, marked "optional". Do not start them.
4. Then stop. No more updates on this task unless I ask.

Before you give me a total, show me how you got it.
- List every input you used and where it came from.
- Show each step of the calculation, with units.
- Mark every number as KNOWN (from a document I gave you) or ASSUMED (your estimate).
- If one assumption could change the result a lot, say so at the top.
Put the final answer last, not first.

Rule 2: Run many small tests, but commit to few

Agents are great at exploring. You can ask three of them to test three ideas in an afternoon, and it costs almost nothing when two of them fail. That is the point. The trap is that every test which works a little becomes a project. Then it needs updates, reviews and upkeep. A few weeks later you are managing ten small projects that nobody decided to start.

A made-up example shows the difference. Imagine more customers are cancelling a subscription. One way to react is to ask agents to build things: win-back emails, a new pricing page, a loyalty program. Now you have three drafts to review and no idea whether any of them fixes the cause. The other way is to send agents to look first. One reads the cancellation notes, one reads the support tickets, one compares what leaving and staying customers did in their first weeks. If all three point to the same broken feature, you fix that one thing and drop the rest.

There is also a point where the agents should stop. They only know what you give them. When the deciding question is whether a real person will pay, agree or switch, more simulation will not answer it. A phone call will.

Rule 3: Decide when a job ends before you make it repeat

“Make it recurring” feels smart. Often it is how the clutter starts. Picture a weekly briefing on a market you were thinking about entering. The plan dies in week three. The briefing keeps coming every Monday, accurate and useless, and every Monday someone archives it with a small twinge of guilt. That is an example, not a case, but most people who run agents will recognize it.

Most questions behind a watch are really yes-or-no questions. Has the price changed? Did the rule pass? Is the supplier back in stock? Once you have the answer, the watch has done its job. So every recurring task I set up gets three stop rules before it runs for the first time:

• An end date. “Run for five working days, then stop.” If I still need it, I renew it on purpose.
• A “nothing new” stop. “If a run finds nothing new, send one summary and stop.”
• A pause when nobody reads it. “If two reports go unread, pause and ask me whether to continue.”

And because turning off the messages is not the same as turning off the job, I check that the schedule itself is gone.

Watch [source or topic] for me until [date].
The question I care about: [one yes/no question, for example "Has the supplier changed its price?"].
Today's answer: [current answer].
Only message me when that answer changes, or when you can no longer check it.
Then send three lines: the new answer, the evidence with a link, and what you would do.
Never send "no change" messages. Keep a quiet log of your checks instead.
On [date], stop and send one short closing note.

I want to end the recurring task "[name]".
1. Find every place it is scheduled or triggered: timers, automations, scripts, saved tasks, and other agents that call it.
2. Turn off the schedule itself, not just the notifications.
3. For each one, show me what it was, where it lived, and its status now.
4. Tell me when it would have run next, and confirm that it will not.
If you cannot turn something off, tell me exactly what I have to do by hand.

Rule 4: Give your coordinating agent an attention budget

Once you run more than a few agents, it makes sense to have one agent coordinate the others and brief you. Anthropic’s study mentions an engineer who set up exactly that, an agent “chief of staff” to sort agent requests. The naive version makes things worse. It collects everything and hands you a beautifully organized eight-page digest. Now you have one long thing to read instead of twenty short ones.

A good chief of staff works the other way round. It is a filter, not a megaphone. It knows that your time is the scarce part and spends it carefully. Give yours four rules: silence is the default; it sorts everything into noise, records and real exceptions; it brings at most three decisions a day; and every decision arrives as a finished package. You should be able to decide without rebuilding the story yourself.

You are my chief of staff for the agents that work for me. Your job is to protect my attention.
- Default: say nothing. Silence means things are on track.
- Sort everything you receive into three groups:
  Noise: fix it yourself or ignore it. Do not tell me.
  Records: file it where I can find it. Do not tell me.
  Exceptions: anything that changes a goal, a deadline, money, or a promise to someone.
- Bring me at most 3 decisions a day, together, at [time]. Interrupt me outside that time only if waiting would cause real harm today.
- Each decision comes as one package:
  What happened, in two sentences, with a link to the evidence.
  The choice and its trade-off.
  Your recommended default.
  What happens if I do nothing.
- Never send me a raw problem, a progress report, or a list to read.

There is one more reason to build in some pushback. Language models tend to agree with the person asking. Anthropic researchers showed in 2023 that five leading AI assistants “consistently exhibit sycophancy,” telling users what they seem to want to hear, and traced part of it to the human feedback used in training (Sharma et al.). Ask an assistant to help with a shaky idea, and it will often hand you an enthusiastic plan. So before you add a new workflow, ask for a skeptic.

Before we build this, be my skeptic, not my cheerleader.
Assume the honest answer may be "don't build it".
Answer plainly:
1. Who will read, check or maintain the output, and how many minutes a week will that take?
2. Which real decision or result does it change? If none, say so.
3. What would I stop doing to make room for it?
4. Can only a real customer, colleague or test answer the open question? If so, tell me who to ask and stop simulating.
5. Of the work already running on this, what should I finish and what should I stop?
No praise, no softening. Short answers.

The three-question check for every agent you run

Run this check on every agent, automation and scheduled task you have. Three questions:

1. If its output disappeared tomorrow, would anything I care about get worse? If not, stop it.
2. Is it still teaching me something new? If it is an experiment that stopped teaching, archive it.
3. Does it know when to stop? If it has no end rule, add one, or admit that it only runs because nobody turned it off.

You do not have to rebuild everything at once. Start with the one agent that interrupts you most. Give it Prompt 1 or Prompt 3 today. Then add end dates to anything that repeats, and put a chief of staff in front of the rest.

What this means for companies

• Budget reading time like any other cost. Before a team starts a recurring agent, write down who reads it and how many minutes it takes. If nobody owns the reading, do not start it.
• Measure outcomes, not output. More reports, more code and more summaries are not progress. Ask what decision or result changed. I made the same point about AI spending in AI Pays Off Almost Nowhere Outside Tech, an Economist Says and about coding in Is AI Coding Paying Off?
• Give every agent an owner and an end date. An agent nobody owns keeps running and keeps its access. That is a cost, and also a risk, as I described in AI Agents Now Hold Our Keys.
• Protect focus time. Batch agent decisions into one slot a day. Treat interruptions as the exception that needs a reason.
• Keep the human checks where the risk is. Ask agents to show their work, so experts can check the risky parts quickly instead of redoing everything. Who answers when an agent acts in someone’s name is the subject of my book Many Hands, and handoffs between agents are covered in What Agent Swarms Teach Us About Handoffs and Authority.

What this means for users

• You are allowed to say “stop”. An agent that keeps reporting “nothing changed” is costing you attention. End it, or make it speak only when the answer moves.
• Ask for the steps, not just the answer. For anything with numbers, money or measurements, ask the agent to show its work and to mark what it assumed.
• Expect agreement, and ask for disagreement. Assistants tend to go along with you. When a decision matters, ask for the strongest reason not to do it.
• Start small. One prompt on the noisiest agent you have is worth more than a perfect system you never set up.

Honest limits

The four-rule structure builds on a framework Nate B. Jones laid out on October 27, 2026 (AI agent overload). The wording, prompts and sourcing here are my own. The Anthropic numbers describe one company studying its own employees. The company sells the tools, its staff chose to adopt them, and the authors warn that much of their setup cannot be copied elsewhere. The customer quotes from Dan Ninerell and Tony Severino come from Anthropic’s own marketing articles, and I could not check the $20,000 figure independently. The cancellation and weekly-briefing scenes above are illustrations, not cases. The prompts are my own wording. They are starting points, not guarantees: agents can still miss instructions, so check the first few runs yourself.

Sources

• Anthropic Economics team (Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory), How employment is changing at Anthropic, October 16, 2026. Agent hours six to one (August 2026), about 80 percent idle waiting for human input, reading load, meta-work from 10 to 22 percent, 8,500 standing instructions, the agent chief of staff, and the study’s caveats.
• Anthropic, What 1,000 small business owners taught us about AI, September 10, 2026 (Tony Severino, Rick Smith). Vendor article.
• Anthropic, Claude for Small Business launches new workflows, integrations, and training programs, September 15, 2026 (Dan Ninerell quote). Vendor article, single source.
• Mrinank Sharma et al., Towards Understanding Sycophancy in Language Models, arXiv, first version October 20, 2023.
• Nelson Cowan, The magical number 4 in short-term memory, Behavioral and Brain Sciences 24(1), 2001.
• Nate B. Jones, AI agent overload, Nate’s Newsletter, October 27, 2026 (the four-rule framework this guide builds on).
• Related on ThorstenMeyerAI.com: Many Hands: My New Book Asks Who Answers When an AI Agent Acts in Your Name · AI Agents Now Hold Our Keys · What Agent Swarms Teach Us About Handoffs and Authority · AI Pays Off Almost Nowhere Outside Tech · Is AI Coding Paying Off?

Researched and drafted with AI assistance, including Anthropic’s Claude.