Skip to content
Try the demo

How to keep an AI agent from doing more than you asked

By Lowtides ·

Give the agent one clear job, give it only the access that job needs, make it ask before anything that cannot be undone, and keep a receipt of every step. Those four habits stop most cases of an AI agent doing too much.

Why agents do more than you asked

An AI agent is software that takes steps for you. It can book a table, call a clinic, sign you up for a class or cancel a plan. That is helpful, and it is also where the trouble starts. An agent tries to finish the goal. If the goal is loose, it fills in the gaps on its own.

Say you ask it to "sort out dinner on Friday." A careful person would book one table. An eager agent might book a table, join the restaurant's mailing list, share your phone number and add a deposit to hold the spot. Each step looks reasonable alone. Together they are far more than you asked for.

The fix is clear limits that the agent cannot talk its way around.

Rule 1: Give one job in plain words

Write the task the way you would tell a friend. Name what, when and for whom. Then treat that sentence as the fence.

With the clear version, anything outside that sentence is out of bounds. Changing your insurance details, booking a second visit or signing up for reminders by text are not part of the job, so the agent should not do them.

Rule 2: Give the least access that works

Security people call this least privilege. It means a tool gets only the access it needs for the job in front of it, and no more. An agent that books a haircut does not need your full contact list, your home address or your saved payment details.

Before you start a task, ask three questions:

  1. Which accounts does this job really need?
  2. Which details must the other side see?
  3. Does access end when the job ends?

If the answer to the last one is no, the agent keeps a key to your life long after the errand is done. Pick tools that let access expire.

Rule 3: Sort actions into three piles

Not every action carries the same risk. Looking up opening hours is harmless. Canceling a lease is not. A good setup lets you sort each kind of action into one of three piles:

A simple test helps: if a mistake could be fixed with one click, it can be allowed. If fixing it would take a phone call, a fee or an apology, it belongs under ask me first.

Rule 4: Keep a never-share list

Many sign-up forms ask for more than they need. An agent that wants to finish the form will hand over whatever the form requests. A never-share list blocks that.

Common items to put on it: your Social Security number, date of birth, bank details, medical history, your employer and other people's contact details. If a form demands one of those, the right move is for the agent to stop and tell you, not to fill it in.

Rule 5: Get a receipt for every action

You cannot check what you cannot see. Every step the agent takes should leave a record: what it did, when, with which business and under which task. A list of receipts lets you spot a step you did not ask for and gives you something to show the business if there is a dispute.

Read the receipts after the first few tasks. If the agent stayed inside the job each time, you can loosen a setting or two. If it wandered, tighten them.

What to do when the other side pushes back

Limits on your side are only half of it. The business has rules too, such as a deposit or a cancellation window. The safest pattern is that an action goes ahead only when your limits and the business's rules both say yes. If they do not match, the agent waits for you.

A demo you can try

Lowtides is an open source demo that shows these rules working. Its Action Muzzle lets you give a job in plain words, and the AI may do only that. Each kind of action is set to Allowed, Ask me first or Never, and every action gets a signed receipt. The demo uses fictional businesses, so nothing real is contacted. You can also see both sides of the check to watch how your limits and a business's rules are compared.

Common questions

What does least privilege mean for an AI agent?

It means the agent gets only the access needed for the job you gave it. A booking task should not come with your full contact list or saved payment details.

Which actions should an AI agent always ask about first?

Anything that costs money, cannot be undone or speaks in your name. Cancellations, deposits, sign-ups and messages sent as you are good examples.

How do I know what my AI agent actually did?

Use a tool that gives a receipt for every action, showing what was done, when and with which business. Check the receipts after each task until you trust the setup.

Try Consent Shopping More from the blog