How it works

A conceptual view of Voice — speech, thinking, and governed models as three boxes, without operator runbooks or engine internals.

Voice is one product with three layers. Callers talk to Talk. Talk asks the Reactive Agent what to say and which action to take. Both use Data Nexus for governed model serving on infrastructure you control.

Three boxes

Caller (phone / website / app) │ ▼ ┌───────┐ │ Talk │ speech in and out, turn-taking └───┬───┘ │ ▼ ┌───────────────┐ │ Reactive Agent │ what to say, which action └───────┬───────┘ │ ▼ Data Nexus (governed models)

Talk — captures speech, plays replies, and handles interrupt. Your website or app talks to Talk; callers never call the thinking layer directly.
Reactive Agent — decides the reply: approved knowledge, an enabled skill, or a handoff. Same Voice product, not a second SKU. See Reactive Agent.
Data Nexus — the private control layer for inference. See Data Nexus overview.

A turn

The caller speaks.
Talk turns that into a turn the agent can use, then plays the reply as speech.
The agent answers from approved knowledge, runs one enabled skill (after a confirm if you require it), or hands off to a person.
If the caller interrupts, Voice stops speaking and listens.

Hang-up and transfer are signals for the channel — they are not spoken as if they were part of the reply.

Channels

Surface	How it connects
Website or mobile app	Your client opens a live session to Talk (microphone audio in, speech audio out). See the Talk API.
Phone	Inbound calls on the customer's phone system reach the same agent. Telephony setup is scoped in a pilot — not published as a public runbook.

Two settings, not one

Production Voice has a voice profile (how it listens and speaks) and an agent (knowledge, skills, persona). Your app passes identifiers Talk and the agent already know. Operators author those settings in the console provided with a pilot — not from this public site.

What you extend

Behaviour on the call comes from knowledge you publish and skills you enable. Skills can be delivered by MinskyAI, by an integrator, or built in-house with the Skill SDK. Enable is the last step.

Deploy, phone-system wiring, and operator manuals stay behind a pilot or architecture review. Next: Reactive Agent or the Talk API.