What you can hand to an agent
Lesson 1 of 22 in the course “AI Agents for Life”. Free, with animations and a quiz.
A chat advises, an agent does
Meet Liza, 19. She studies design, works at a café and never has enough time. A regular AI chat answers her with text, and she does everything else herself. An agent is AI that carries out the steps on its own: it searches, compares, fills in forms and reminds you. Anthropic, the company behind Claude, puts it like this: an agent decides by itself which steps to take and which tools to use. How an agent works inside is covered in the lesson “Agents: AI that takes action” in the AI from Zero course. Here we’ll build our own.
What agents can already do
Almost everything agents do well falls into four buckets: find and compare, clear a pile, keep watch on a schedule and do something on a website. An example from the launch of ChatGPT agent in July 2025: “plan and buy ingredients to make Japanese breakfast for four.” The agent finds a recipe, works out the ingredients and goes shopping. Before important actions, such as a purchase, it asks for permission. You’ll see that rule many more times.
What to keep for yourself for now
The rule is simple: hand an agent things where a mistake is easy to spot and fix. A bad shopping list is no big deal, you’ll grab what’s missing. A treatment decision, a big money transfer or a message you can’t take back is a different story. An agent can be wrong in a very confident tone, and you are the one who lives with the result. Never give passwords or SMS codes to anyone, AI included: a confirmation code is the key to your account.
Why long tasks break
Say an agent gets each step right 95 times out of 100. Sounds reliable. But the odds multiply: 0.95 × 0.95 × 0.95… After 10 steps the chance that nothing went wrong is about 60%, after 20 steps about 36%. This is plain probability math and it applies to any agent. The takeaway: cut a big task into short pieces and check the result after each one.
Start with the simplest thing
Anthropic’s engineers advise finding the simplest solution and adding complexity only when you truly need it. Often one good question in a chat is enough. A brief with steps is for tasks with several parts. An agent with tools is for when it has to go into your email or websites itself. An agent on a schedule is for tasks that repeat and should run without you. Climb the ladder only when you hit a step.
Try it: a week audit
Write down 10 things your last week went on. Ask three questions about each: does it repeat? is it boring? is a mistake easy to fix? Three “yes” answers mean it’s a job for an agent. Liza found three: email, paint discounts and her shift schedule. Her term paper and messages to her mom she kept for herself. Too lazy to write the list? Copy the brief below into ChatGPT, Claude or Gemini: the AI will ask the questions and build the table.
Check yourself
- Is this a job for an agent or a regular chat?
Correct answer: An agent: it repeats and should run by itself. The task repeats every day and should arrive without you asking. That is exactly why agents got schedules and access to your data. - What’s the chance all 10 steps go right?
Correct answer: About 60%. The odds multiply: 0.95 to the 10th power ≈ 0.6. That’s why you cut a long task into pieces and check the result along the way. - What do you tell your friend?
Correct answer: I decide on big money myself; the agent can gather and compare options. Keep decisions with costly, irreversible mistakes for yourself. An agent is great at gathering and comparing options, and you make the choice. - Where is it best to start?
Correct answer: With one good chat question; bring in an agent when the task repeats and has many steps. Anthropic’s engineers advise the simplest solution first and more complexity only when needed. Every extra step is a new place for a mistake.