Tool use
The broader ability of an LLM to use external tools — search, code execution, a calculator, databases, APIs — to do things it can't on its own. Function calling is the mechanism; tool use is the concept and behavior.
Why: a language model is bad at exact arithmetic, has no fresh information, and can't act in the world. Tool use gives it "hands": ask for a web search for current facts, run Python for a calculation, query a database, call an API to book something. The model reasons about which tool is needed, calls it (via function calling), reads the result, and continues — often in a loop (reason → act → observe), which is the core of an agent. Win: dramatically expanded ability and reliability (the tool does what the model is bad at). This is the basis for agentic systems, the ReAct pattern, and MCP. Distinct from pure text generation by the model affecting and fetching from the outside world. Related to function calling and ReAct agent.