What is the difference between a function-calling model and a true reasoning agent?