points
Basically it can run these mini programs where each input might be another toolcall, so it can run without waiting for whole LLM response and ready the parameters async.