You keep picking the recommended option. You'd be better off doing something else.

Work with agents for a few weeks and you end up clicking the same button all day.

It starts innocently. An agent puts a choice in front of you: which wording, which order, which of two routes. It tells you what it would pick. You look, think that makes sense, and choose the recommendation. An hour later the next question arrives. By the end of the day you've done the same thing thirty times, and the agents have waited thirty times.

I recognised it in myself. I'd estimate I chose the recommended option about nine times out of ten. One afternoon I asked myself the question that goes with that: if I do this all day, what's a better choice?

In this piece I explain why the pattern is so stubborn, where the line sits between what you let go and what you keep, and how to tackle it yourself in four steps. It's a way of working and a way of thinking, not a product. You can start today with a notepad.

The bottleneck isn't the agent

When autonomous agents get stuck, we usually look for the cause in the agent: too little knowledge, too few tools, a weak model. Sometimes that plays a part. But after working with a team of agents for a while, you see something else. The work piles up because a human has to choose.

Every question to you stops the work. The work stays stopped until you answer. And the nasty part is that you only notice this once there are many agents. One agent asking now and then is pleasant: you keep an overview. Ten agents each asking now and then make you the narrowest point in the whole process.

It's the same lesson as in any other process. You don't move a bottleneck by making everything else faster. I could make the agents as good as I liked; throughput was set by how fast I clicked buttons.

A choice is only information if it changes something

Here's the core. If you pick the recommendation nine times out of ten, your choice is no longer information. It doesn't change what happens. The agent would have done the same thing an hour earlier, without you needing to be there.

That's not a reproach to yourself. It shows the recommendations are good, which is exactly what you want from an agent. But it also means you're paying, in waiting time and attention, for a decision you aren't making. You're adding a tick mark.

So the better question isn't "how do I get better recommendations?" but "which choices must I make myself, and which not?"

The line sits at danger, not at importance

My first instinct was to draw the line at what's important. That doesn't work, because almost everything feels important when it's yours. A headline on an article is important. The time of a post is important. Yet it would make no difference if the agent made that choice.

The line that does work sits at danger. I mean it like this: a choice is mine if it has a different outcome that's hard to undo, or if it touches something I don't hand over. That comes down to a few fixed kinds:

• Money: paying for something, taking out a subscription, naming a price.
• My name going out, for the first time: a new kind of statement, a first message to someone.
• People: what a real person receives or notices, and anything involving personal data.
• Legal matters: contracts, terms, liability, what I do or don't offer.
• What can't be undone: deleting something, anything that stays in a public history.
• Access and security: keys, permissions, a new connection to the outside.
• The big direction: a new product, a new service, anything that takes a lot of work or money.

Everything else is a difference of nuance. Word choice, order, formatting, a time within the day, the choice between two equivalent routes. A nuance has no different outcome; it's another road to the same place. I don't need to choose there.

If you doubt whether something falls in the list, it does. But doubt about a nuance is not doubt about the list. That distinction is the whole job.

Letting go becomes safe when undoing is one action

The objection is obvious: what if the agent gets it wrong? It can. That's why every decision the agent takes itself comes with a second line: how I undo it.

An agent that chooses for itself writes in its log what it chose, why, and how to reverse it. For example: "chose the short headline because the long one broke the line; reverse by putting the old headline back in the file". That's one sentence, and that's what it's about. I don't need to know in advance whether the choice is good. I need to be able to see afterwards what was chosen, and to turn it around if it's wrong.

This flips the idea of control. You no longer check before every step; you look back over a row of decisions and step in where needed. That's far less work, and it's more honest: you look at what really happened, not at what might happen.

A person manages the rule; the agent doesn't invent it

This is where it can go wrong if you do it the other way round. You can tell an agent "use your own judgement on small things". Then everything depends on what the agent considers small, and that shifts with the day and with the model.

Better to record decision rules as something you manage: a list in plain language that says, per rule, when it applies, what the agent may then do, and how you reverse it. The agent may propose a rule. Only you switch it on. A proposal could sound like this: "over the past weeks you chose the recommendation five times for this kind of question; do you want me to do this myself from now on?"

That makes autonomy not a switch but a list of rules that grows with your trust. You start with three rules. After a month there are ten. After half a year you know exactly where your trust lies and where it doesn't, and it's on paper.

So don't write down what the agent must do, but what you want to keep deciding yourself. The rest follows from that.

The questions that remain get better

There's a side effect I didn't expect. When the ordinary questions disappear, what remains is heavier and more meaningful. You only get the choices that something hangs on.

And you can ask those questions more cleverly. A question you put forward should bring its follow-up question along. If an agent asks "may I send this?", it should already know what it still needs from you after a yes, and after a no. Otherwise you give a yes, and the work gets stuck again a step later because a detail was missing.

Best of all is when follow-up questions only appear when they're relevant: you choose, and only then does the next question open up. So you don't get a form of twenty questions, but a conversation that follows what you say.

What you can do today

You don't need a studio or special tools for this. Four steps, over about four weeks:

1. Count for a week how often you pick the recommended option, per kind of question. A tally sheet will do. You'll see that some kinds almost always end the same way, and others never do.
2. For the kinds where that's almost always so, write one rule: "the agent may do this itself, and this is how I undo it". One sentence per rule. Start small, with two or three.
3. Write down the list of what always stays with you: money, your name going out, people, legal matters, what can't be undone. That list outranks every other rule. If a rule clashes with the list, the list wins.
4. Look back after a month: what was reversed? Where something came back, tighten the rule. Where nothing came back, it may be wider.

That last part is the one people skip, and it's the most important. Without looking back it's a gamble. With looking back it's a learning process in which you and your agents establish together how much room there is.

What it delivers

I mainly see two things. The work carries on while I'm not there, because the agent no longer waits for a click that would always be the same anyway. And my attention goes to the places where it counts.

There's a third thing, and it's subtler. By writing down what I want to decide myself, I know it better too. The list of what stays with me is a kind of statement of what I want to be responsible for. That's something you otherwise never get so clear for yourself.

If you click the same button all day, you're no longer a decision-maker but a stamp. Then it's time to decide where you do want to decide, and to let go of the rest with a route back.

In my book From Zero to AI Agents That Do The Work I work out how to set up agents so they can work within your limits, with assignments you can do the same evening: https://florisknol.com/books/ai-agents-that-do-the-work/

All pieces