Tool use, often called function calling, is the mechanism that lets a language model do more than produce text: it can decide to call an external function, run a search, query a database, or hit an API, and then use the result to continue. Instead of guessing an answer, the model emits a structured request (“call get_weather with city=London”), the surrounding software runs that function, and the model receives the output to reason over. This is the technical step that turns a model into an agent.
The foundational idea appears in the 2022 paper “ReAct: Synergizing Reasoning and Acting in Language Models” (Yao et al., arXiv 2210.03629), which interleaved step-by-step reasoning with concrete actions so a model could both think and act inside one loop. The abstract describes prompting models “to generate both reasoning traces and task-specific actions in an interleaved manner.” Model providers later standardized this into a formal “function calling” interface, where developers describe the available tools in a schema and the model returns a structured call when it needs one.
The distinction from the related ai-agents entry is worth keeping clear: an agent is the overall system that plans and pursues a goal; tool use is the specific capability that lets that system reach outside itself. Without tool use, a model can only describe what it would do. With it, the model can actually check a calendar, send a query, or run code.
Why business readers should care: function calling is the single feature that moves AI from giving advice to performing work inside your systems. It is also where most of the real risk lives. A model that can call functions can take wrong actions, leak data through a tool, or be manipulated into misusing a tool, which is why permissioning and auditing of available tools belongs in any serious deployment plan.