[minimaxir]

The LLM is what decides which endpoint/tool to call (or none at all) in response to the user input.

The original 2022 ReACT paper is still best explainer: https://arxiv.org/abs/2210.03629