Inject Agent

Voice Agent

The InjectAgentMessage message is a JSON message you can send to immediately trigger an agent statement.

InjectAgentMessage lets your server put words in the agent's mouth mid-conversation. The optional behavior field controls how the message interacts with any ongoing user or agent turn: wait for silence, queue within the current turn, or interrupt the current turn.

• default: the agent speaks only if neither the user nor the agent is mid-turn. If a turn is in progress, the server rejects the request and replies with InjectionRefused. This matches the original InjectAgentMessage behavior.
• queue: the server appends the message after any ConversationText that is already queued, without interrupting the current agent turn or the in-flight think response. If nothing is queued, the agent speaks the message immediately.
• interrupt: the agent immediately speaks.
  • If the agent was already speaking, it interrupts any current speech and replaces it with the new message.
  • If the user is speaking, the agent interrupts with the new message. However, the user's continued speech triggers UserStartedSpeaking, which quickly interrupts the agent, ensuring the conversation continues forward.

JSON

{
  "type": "InjectAgentMessage",
  "behavior": "default",
  "message": "Are you still on the line?"
}

JSON

{
  "type": "InjectAgentMessage",
  "behavior": "queue",
  "message": "Thanks for your patience, the system is still loading."
}

JSON

{
  "type": "InjectAgentMessage",
  "behavior": "interrupt",
  "message": "Sorry to cut in — let me correct that."
}

The server sends an AgentAudioDone message after the last InjectAgentMessage is spoken.

JSON

{
  "type": "AgentAudioDone"
}

When behavior is default or queue and the request arrives while the user is mid-turn, the server ignores the request and replies with InjectionRefused. If the agent is mid-turn, the server returns InjectionRefused only when behavior is default. The interrupt behavior is never refused: the agent speaks even while the user or agent is mid-turn.

JSON

{
  "type": "InjectionRefused"
}

Pick the behavior value that matches what you want to happen to the current turn.

Use default when the injection is only appropriate during silence and you would rather drop it than step on the conversation.

• Prompting the user to continue if they have been silent for a while ("Are you still on the line?").
• Optional nudges where being ignored is acceptable if the caller is already talking.

Use queue when you want the agent to say something after whatever it is currently saying, without stepping on it. This is the right choice for filler during function calls, where the model has already emitted pre-function narration and you want to extend it rather than replace it.

• Filling silence during a long-running function call ("One moment while I pull that up") without cutting off the model's current response.
• Chaining a follow-up sentence after the agent finishes its current turn ("...and I've also emailed you a copy.").
• Streaming progress updates from a backend job so the caller hears them in order.

Use interrupt when the new message must take over immediately, even if it means cutting off the agent's current speech. Because the agent speaks regardless of whether the user is talking, reserve this for time-sensitive corrections and overrides.

• Correcting the agent mid-sentence when it is saying something wrong or outdated.
• Delivering an urgent update that should replace whatever the agent is currently saying.
• Overriding the current turn from a supervisor or backend system that has higher-priority information.