This page was translated from Dutch with AI.

Tell your agent what not to do as well

You give your agent a clear task and assume it will understand the rest itself. I understand that reflex. Yet in workshops I often tell the story of someone who gave an AI agent access to a credit card. The instruction was clear: book the cheapest flight to Bali. The agent could not find any standard seats left and booked a business class ticket with an expensive airline. Technically, that was the cheapest available option. A few thousand euros later, the owner understood what had gone wrong.

Feel free to copy the prompt at the bottom of this article, so you can immediately write the boundaries for your own agent.

Cartoon
Vertel je agent ook wat hij niet mag doen: grenzen in de instructies van je AI-agent

The agent did exactly what was asked

The painful part of this story is that the agent actually did nothing wrong. It was given a goal, it had the means and it carried out the task. What was missing, was a framework. No one had said that business class was not an option, and there was no maximum budget anywhere. Nor had anyone asked to consult first before it paid.

In the workshop I put it like this: you determine what it can and cannot do. If you do not give it that framework, it will do what it can do. And an agent with access to your mailbox and your calendar, and to your payment methods, can do a lot. You are probably already thinking of a task for which you would like to use such an agent. Perhaps you are wondering what could go wrong there.

Frequently asked questions

Which boundaries should I put in the instructions for my AI agent?

I would start with the actions you cannot reverse, such as paying or deleting, or sending something in your name. Have the agent make a proposal there first and wait for your agreement. Actually, also add a maximum amount and state which folders and tools it is allowed to use.

How much access do I give an AI agent?

Just as much as it needs for that one task. An agent that manages your schedule does not need access to your accounting. The smaller the playing field, the smaller the damage if something goes wrong.

What should an AI agent do when in doubt?

I would put in its instructions: when in doubt, ask me. To you many boundaries are obvious, but to an agent a boundary only exists if it is written down. With that one sentence it stops and asks first.

Where can I practise building an agent with good boundaries?

In AI agents: build your own digital colleague you practise hands-on, with your own work, how to build an agent and write its instructions. Claude AI Power User: automate your workflows also fits with that. You can find all dates and registration at Workshops & Events.

Keep reading for free

Register free with your e-mail address and read on. Your account opens the free articles in the learning space; the rest is for subscribers, from Learn Plus up.

Already registered? No account yet?

No comments yetRead the comments

Comments

Members comment here and read each other's comments. Register for free or sign in to join the conversation.

Register for free or sign in

No comments yet. What do you take away?

    QR code of this page

    coachsteff.live/en/wat-je-agent-niet-mag

    Scan with your phone to open the page.