You type a question into a chatbot and it answers. Once. If the answer needs something the model does not have (today's price, your calendar, the contents of a file), it either guesses or tells you it cannot. An agent is the same model with two additions: tools it can call to get or change things, and a loop that lets it call them repeatedly until the task is done.
| Chatbot | Agent | |
|---|---|---|
| You ask | "Book me a table for four on Friday near the office" | "Book me a table for four on Friday near the office" |
| What happens | It writes a polite reply suggesting you check a booking site | It searches nearby restaurants, checks Friday availability, picks one that fits, books it, and tells you the confirmation |
| Turns | 1 | As many as it takes: 4 to 8 here |
| What it needed | Nothing | A search tool, a booking tool, a rule for when to stop |
| What it cost | One short exchange | Every turn re-sends everything so far (lesson 2) |
One run, narrated
The booking agent, step by step
- 1
Read
The agent sees your request, its instructions, and the list of tools it is allowed to use. Nothing has happened yet.
- 2
Decide
It decides the first useful action is to search for restaurants near the office address. It writes that as a tool call, not as prose to you.
- 3
Act
The search tool runs and returns a list of eight places. That list is appended to the conversation.
- 4
Read again
Now the agent sees the request, the tools, its own search call, and the eight results. It decides to check Friday availability at the top three.
- 5
Round and round
Three availability checks, three results, one booking call, one confirmation. Each time, the whole history so far is what it reads.
- 6
Stop
It has a confirmation number, so it writes you a final message and the loop ends. Without a stop rule it could keep checking restaurants forever.
Knowledge check
What are the two things an agent has that a plain chatbot does not?