How AI agents work: an AI that can only write, and still gets things done
Tool calls: how an AI agent acts when its model can only write
Ask an AI assistant if it will rain tomorrow. It answers with the real weather forecast. But the AI model inside it can only write text, and it cannot go on the internet. So who got the forecast?
-
1
Saturday is picnic day at the park with your friends Miva and Rono. So you ask your assistant app: will it rain?
-
2
The app sends your question to the model, the program that writes the answers. It has no eyes on the sky and no hands for a button. All it can ever do is write text.
-
3
So it writes something else: a short request in a form the app can read, with the tool, the place and the day. Then it stops and waits.
-
4
The app reads that request and does the work itself. It runs a tool, a small program with one job, here the weather. The request is a tool call.
-
5
The tool answers: rain from 2 pm. The app pastes this tool result under your question and sends it all back to the model.
-
6
The model reads the result like any other text and writes your answer. It never touched the weather: it asked, and the app did the rest. That pair is the heart of every AI agent. Yet you never told the model a weather tool exists.
-
Quick check
Lights out.
Show the answer
The app. The model only writes a request for the light tool, and the app runs it.
-
Go further
Two names you may meet.