How AI agents work: an AI that can only write, and still gets things done
Tool descriptions: how an AI agent picks the right tool
An AI agent can have dozens of tools: weather, calendar, email and more. The AI model never sees the code of any tool. So how does it know which tools exist, and which one to use?
-
1
Look at everything the app sends the model with your question. Above your words sits a list you never typed: the tools the app can run.
-
2
Each tool gets a name, a few plain words on what it does, and the blanks a request must fill. This is the tool description.
-
3
The model reads that list the way it reads any text. Your question is about rain, and that fits the words of the weather tool best, so it asks for that one, and fills the blanks with the park and Saturday.
-
4
The model never sees the weather program itself. It knows the tool only from what the app writes about it, so those words decide.
-
5
Give the weather tool a vague line, and the model may skip it, or pick the wrong tool. That is why people who build agents write these lines with great care.
-
6
Now you ask for more: if it rains on Saturday, move the picnic to Sunday and tell Miva and Rono. That takes three tools, one after the other.
-
Quick check
A trip to Rome.
Show the answer
Hotels. The model compares what you ask with what each description says, and sleeping fits rooms, not flights.
-
Go further
Two more things about the list.