From Chat to Agent

A chat assistant writes. Give it tools and it can do things: look up the weather, search your email, add a calendar entry, book a table. A tool is a request the app makes to another service on the model’s behalf, much as a weather app asks a forecasting service for tomorrow’s rain. And the request is made with the access you granted: your account, your calendar, your email.

Here is the mechanism to hold on to. The model can’t press anything itself. It writes a proposed request (“look up Bristol”, “send this email”), and the app checks it is allowed, runs it and hands back the result. So the model’s proposed call is not proof that anything happened. Only the result, and for a real change the state of things afterwards, tells you that.