How AI agents work: an AI that can only write, and still gets things done
The agent loop: how an AI agent keeps going until the job is done
One request can need three tools in a row: check the weather, read the calendar, send a message. But the AI model asks for one tool at a time, and then it waits. So what keeps the AI agent working until the job is done?
-
1
Round one starts as before: the model writes a request for the weather. The app runs it and pastes back the result: rain from 2 pm.
-
2
Round two. The model keeps nothing between runs, so the app sends it everything again, that result included. The model reads "rain" and asks the calendar tool to move the picnic to Sunday. The app runs that one too.
-
3
Round three: a request to the messages tool, for Miva and Rono. Sent.
-
4
In round four, the model writes only words for you, with no request in them. For the app, that is the signal to stop: it shows you the answer.
-
5
Draw it as a circle: the model writes, the app runs, the result goes back. This is the agent loop. A model and an app that run it together are what people call an AI agent.
-
6
On Miva's and Rono's phones, the message is already there. You asked for it, true, but nobody showed you the words before they went out. So who decides what an agent is allowed to do?
-
Quick check
Count the runs.
Show the answer
Three. Once for each request, and once more for the answer with no request.
-
Go further
One name from the research.