Agents

The worker is a participant, running next to the media plane

It joins the room the way a person does and is dispatched by the control plane when the call starts. Running beside the media server rather than across the internet from it is the difference between a reply that lands in the conversation and one that arrives after it moved on.

Listen, decide, speak, and record what was said

Speech in, your logic in the middle, speech out — and a transcript written as it goes rather than reconstructed afterwards. The transcript is a product of the call, not a job queued behind it.

Listen

streamed from the room, not polled

Speak

your models, your keys

Record

transcript and history, per organisation

A face on the call, with consent recorded before the first frame

Enrolment is a separate, long-running job on its own hardware, deliberately not in the path of a call joining. What a person consented to is written down at enrolment rather than asserted later.