Tool use is the mechanism that lets an AI move from "answering questions" to "taking action." The model is given a typed list of functions it can call (look up a patient, check availability, refund a charge), and it decides, turn by turn, which to call and with which arguments.
A production-grade tool use system has three properties: the tools are typed (the schema is enforced), they are scoped (the model can only call tools it has permission to), and they are replayable (you can re-run the same conversation in staging and verify the same calls fire).
Without these, tool use is unsafe. The model can invent tool calls, fire them at unexpected times, or pass garbage arguments. The well-known failure modes are not theoretical, they are the reason most public agent demos do not survive contact with a real CRM.
Vorel ships agents where every tool call is compiled from a typed SDK, scoped to the customer, and replayable against staging fixtures.