Tool calling

Tool calling is the capability that lets a language model reach beyond its own text output and do something — run a search, write to a database, send an email, execute code, or hit a third-party API. Without it, a model can only describe what it would do; with it, an AI agent actually does it and reads back the result.

Under the hood, the agent runtime gives the model a menu of tools, each described by a name and a schema of expected inputs. When the model decides a tool is needed, it produces a structured request, the runtime runs that tool, and the output is fed back into the conversation. This request-execute-observe cycle is the heartbeat of an autonomous agent, and the structured request itself is produced through function calling.

Why it matters: tool calling is what turns a chatbot into a worker. A model asked for today's exchange rate can only guess from stale training data — but an agent with a currency tool calls it and answers from the real number. Standardizing how tools are exposed is now common enough to have its own protocol; see the Model Context Protocol for a portable way to connect agents to tools across vendors.

Concrete example: a support agent receives "Where is order #4471?" It calls lookup_order(id=4471), the runtime queries the shipping API, and the observation comes back as a tracking status. The agent then phrases a natural reply grounded in that live status — no invented dates, no hallucinated carrier.