Voice agent

Voice adds one hard constraint text does not have: the reply has to start before the silence becomes uncomfortable. In text a two-second wait reads as thinking. In audio it reads as a dropped call.

That budget has to cover transcribing the speech, retrieving the answer, generating it and synthesising it back to audio. Architectures that work well for chat (a chain of separate calls) usually blow it, which is why real-time voice models stream the whole thing at once.

The rest is the same product: same knowledge base, same guardrails, same hand-off. Only the latency ceiling and the failure modes (background noise, cross-talk, unfamiliar accents) are specific to audio.