Function calling

The mechanism is a contract. Each function is declared with a name, a description, and a schema for its parameters. The model, given the conversation, decides whether to call one and emits a structured invocation with arguments. The platform executes it, returns the result, and the model continues with that result in context. The model never executes anything itself — it only requests.

Description quality determines behaviour far more than most teams expect. The model chooses a function based on its description, so vague or overlapping descriptions produce wrong selections, and a parameter without a clear type and explanation produces malformed arguments. Fewer, well-described functions consistently outperform many similar ones.

Voice adds a timing constraint that text does not have. A function that takes several seconds leaves the caller listening to nothing, which reads as a dropped call. The standard mitigations are a filler phrase spoken while the call is in flight, an aggressive timeout, and a defined utterance for the timeout case so the agent says something coherent rather than stalling.

Functions are also a security boundary. Arguments are model-generated text and must be validated as untrusted input, permissions must be scoped to what the agent legitimately needs, and any function with side effects needs idempotency protection — a retried call must not book the appointment twice.