All learning paths Learn See Try Make What an agent actually does What this lesson is about An agent picks a tool, uses it, reads the result, and picks again. Everything else is decoration.
Your shopkeeper cousin has been sold an agent for his billing and asks what it actually does. What do you say?
A It reads what you type and writes a reply, the same as a chat assistant. B It works out the whole plan first, then carries the plan out from start to end. C The loop is four turns: pick a tool, use it, read the result, decide again.
Check my recall Why this matters People are sold agents as magic and then cannot say what went wrong. The Hugging Face course describes one loop: think, act with a tool, read the result, think again. Once you can see those turns, you can point at the one that failed.
Describe one agent run as the four turns of its loop, and say where a mistake sits.
People are sold agents as magic and then cannot say what went wrong. The Hugging Face course describes one loop: think, act with a tool, read the result, think again. Once you can see those turns, you can point at the one that failed.
The loop is four turns: pick a tool, use it, read the result, decide again. An agent's mistakes are tool mistakes: wrong tool, wrong input, or a bad result believed. An agent cannot know what nobody gave it, and cannot undo what it already did. Task Watch an agent asked to find the cheapest morning train and write down what it did.
Weak approach Writes that the agent understood the request, searched properly, and then made a mistake at the booking. Nothing names the search it ran or the list it read.
Stronger approach Writes four turns: searched trains, read a list of evening seats, decided the first was cheapest, booked it.
Why the stronger approach works The weak note has nothing anybody can check. The strong one names each tool call, so the wrong input is visible.
Try a changed situation An agent sends a customer a price that is six months old. Its wording was perfect. Where do you look for the mistake?
Make something yourself Watch one agent run. Draw it or number it as four turns. Name the tool it picked, the input it gave, the result it read, and the next decision.
Lesson 1 of 9 on Running and checking AI agents next Write a task an agent can finish Part 1 ends in a work sample you can send
Read the primary or official source: Hugging Face Agents course · the agent loop