When should a voice agent speak?

A voice assistant that runs background tasks needs to decide more than what to say. It must also decide whether an update deserves the user's attention, and whether now is a suitable moment.

Separate work from conversation

Long-running tasks should not hold up the foreground conversation. Executors report progress and results to the runtime; they do not speak directly. The user can ask another question, add a constraint or cancel while work continues.

An event is not automatically an interruption

A completed step, a camera observation and a request for permission have different value. Routine activity can stay in the background. A meaningful result or a decision that needs the user may deserve a response.

Nova separates background suggestions from user-awaited results. Surrogate evaluates whether a background suggestion is worth surfacing. Work the user is waiting for does not depend on that selection step to deliver its result.

Keep one speaking path

Floor coordinates speech. If the user is talking, Nova waits. A higher-priority response may interrupt Nova's own playback, but an executor cannot bypass this control and start speaking by itself.

This separation also keeps priorities under host control: an external tool cannot make itself urgent by putting instructions in its output.

Preserve control as capabilities grow

Documents, personal memory and camera observations can help Nova answer a question or propose a follow-up. They are evidence, not authorization. Actions still pass through the host's permission and project-confirmation rules.

New models and executors can change how work is performed. The useful boundary remains the same: keep conversation responsive, report information selectively, and leave consequential decisions with the user.

See the runtime overview, context view and design constraints.