Models calling external tools instead of guessing.
LLMs struggle with reliable math, can’t access live data, can’t take actions in other systems. Tool use closes that gap. The model generates a structured call (function name + arguments) instead of a natural language answer. A runtime executes the call and feeds the result back. The model never runs anything itself. It decides what to call and why. The runtime handles execution.
The modern implementation. The model outputs structured data matching a defined format, not free-text that needs parsing. More reliable than older ReAct-style text extraction, but the conceptual loop is the same: the model requests an action, something else executes it, the result comes back as context.