How a model requests a structured, directly executable action instead of just describing one in plain text.
An LLM combined with tools it can call and memory of what it has already done — why that combination can act, not just talk.
Why the first retrieval pass often gets the order wrong, and how a second, smarter pass fixes it.
