The brief: how to give an agent a task
Lesson 3 of 22 in the course “AI Agents for Life”. Free, with animations and a quiz.
A brief in 5 parts
An agent can’t read minds. Anything missing from the task, it guesses, and it often guesses wrong. That’s why you give an agent a brief. A brief is a short task in five parts. Goal: what you want to get. Data: what it should know about you and the task. Report: what shape the result should take. Where to ask you. And what never to do. You’ll see this card in every lesson of the course: it’s how any agent is put together.
A bad and a good request
“Find me a flat” isn’t a task, it’s a riddle. Where? For how much? Who are you living with? Is a cat OK? The agent fills the gaps itself and brings five flats that don’t suit you. Now a brief: up to $600 a month, no more than 20 minutes from campus, cat allowed, 5 options as a table, don’t message landlords yourself. Same agent, same model, completely different result. Half of what looks like agent “stupidity” is just a task that left things out.
Stop rules
A stop rule is a line that says “if… stop and ask.” You need one wherever a mistake costs money or nerves: a price over your limit, a message to a real person, missing data, payments and sign-ups. Write them right into the brief, short and specific. Remember the example from the ChatGPT agent launch? It asks before a purchase by itself. But only you know the rules of your life: nobody else will write your $600 limit in for you.
“Done” has to be checkable
“Make a good meal plan”: the agent doesn’t know what “good” means to you, and you can’t check whether it did the job. Turn it into numbers and conditions: 7 dinners, each under 30 minutes to cook, the whole week under $40, no nuts. Now you can see in a minute whether it’s done. A good rule: if you can’t say how to check the result, the agent won’t know when to stop either.
The trick: let the agent ask first
Don’t want to think through every detail yourself? Add one line at the end of the brief: “Before you start, ask me the questions without which the result won’t be accurate.” The agent finds the gaps itself: how many people, any allergies, how much time to cook. You answer in a few words, and only then does it get to work. It’s the cheapest way to raise quality: one line, far fewer mistakes.
Try it: a chef agent
Take a photo of your fridge and shelves, upload it to your assistant and paste the brief below. The agent will work out what you have, ask its questions and build a weekly menu with a list of what to buy. That’s how Liza stopped throwing away forgotten food. Look at the card: goal, data, report, a stop rule on budget and a ban on ordering anything. Want to practise risk-free? Open the “Build an agent” trainer on the course page.
Check yourself
- What is the task missing most?
Correct answer: A budget and a stop rule “over the limit, ask me”. Without a budget the agent can pick anything, and without a stop rule it may go all the way to payment. A number and “stop and ask” fix the main thing. - Which definition of done is better?
Correct answer: A 14-day plan, 45 minutes a day, days off on shift days. Numbers and conditions can be checked in a minute, and the agent knows when the job is finished. - What’s the fastest way to fill gaps in a brief?
Correct answer: Ask the agent to first ask the questions it needs for an accurate result. One line makes the agent find what it’s missing. You answer briefly, and the work starts from the right data. - Which stop rule protects Liza?
Correct answer: “Before any message to a landlord, show me the text and wait for my yes”. A message to a real person can’t be undone. “Show me and wait for my yes” leaves the last word with Liza.
← Assistants in 2026: who’s who Memory: an agent that knows you →