​
Use Conversation and Agent Nodes

​
Global prompt and current stage

The workflow global system prompt establishes universal rules. Every active workflow Agent receives it. A node then adds its current-stage instruction, which prevents the Agent from handling another step prematurely.

​
Conversation

Conversation does not create a new LiveKit Agent. It updates the current active Agent’s instructions and completion policy.
The first two stages use the same main Agent. The specialist node creates a new scoped Agent.

​
Direct Agent

An Agent node starts a fresh scoped Agent. Use it when the next stage needs its own prompt, model, voice, transcriber, tools, files, or resolution outputs. It remains the active Agent after Say, Logic, Function, integration, and Task nodes until a later Agent or Subagent handoff. Configure:
  • Instructions and completion contract
  • Model provider and model
  • TTS provider, voice, and language
  • STT provider, model, and language
  • Organization tools and files
  • Avatar scope: keep the current avatar, disable visuals, or switch to an organization avatar
  • Timeout/recovery routing
An Agent node can supply its own model, voice, and transcriber objects. Those values apply from the handoff until another Agent or Subagent replaces the active Agent. Use the same provider IDs, credential references, vocabulary, fallback-provider ordering, voice behavior, and audio rules as an Assistant; see Configure Your Assistant. If an Agent node does not override a setting, it inherits the workflow’s active/default assistant configuration. Keep provider changes inside an Agent node when the change is specific to that stage, such as switching to a language-specialized STT/TTS pair after language routing.

​
Subagent

Subagent loads an existing Assistant and applies its optional overrides. Existing saved Subagent workflows remain supported, but use a direct Agent in new workflows when you want the workflow to own the configuration. Subagent handoffs support the same Avatar tab and scope behavior as direct Agent handoffs.

​
Start-node speaking behavior

The Start node may set data.start_speaking_plan. It overrides the inherited assistant plan for that workflow session. Use wait_seconds, smart_endpointing_mode (off or livekit), and the endpointing/turn-detection fields described in assistant speaking behavior. A later Agent handoff can apply its own configuration; otherwise the Start-node plan remains in effect.