letting the LLM decide what action to take

splitting a big task into several smaller subtasks with JSON for the LLM

outputting JSON (instead of text) as a response to a prompt

using the output from an LLM for another prompt (and so on)

making queries to systems with natural language