Skip to main content

Design patterns

Recurring ways to arrange agents, tasks, and tools in a voice AI app.

Overview

A pattern describes how to arrange agents, tasks, and tools to solve a class of problem. The topics in this section document patterns that need more than one construct working together, with complete examples you can adapt.

Patterns aren't mutually exclusive. A single app often combines several, and different phases of one conversation can use different patterns.

Choosing a pattern

Both patterns in this section delegate work. They differ in whether the conversation waits for the result.

PatternConversation flowUse cases
SupervisorWaits. A specialist task takes temporary control and returns a typed result before the conversation continues.Data collection, verification, and other discrete operations with a structured result.
Subagent delegationContinues. A slower model works in the background and the result arrives in a later turn.Open-ended analysis, multi-step research, and planning.

For a comparison across every structural construct, including handoffs and task groups, see the table in Workflows.

Additional patterns

The LiveKit blog covers additional patterns.