To make your agent initiate a conversation before the user says something, you can use one of the following methods: 1. Using

session.say()

This method allows the agent to speak a predefined message. For example:

agent.start(ctx.room)
await agent.say("Hey, how can I help you today?", allow_interruptions=True)

This triggers the configured Text-to-Speech (TTS) to synthesize and play the message to the user. You can also provide pre-synthesized audio to skip the TTS step. 2. Using

session.generate_reply()

This method allows the agent to generate dynamic responses. You can implement this in the

on_enter()

method of your Agent class:

class IntroAgent(Agent):
    async def on_enter(self):
        self.session.generate_reply(instructions="Say hello to the user")

Alternatively, you can initiate the conversation at the session level:

await session.generate_reply(
    instructions="Greet the user and offer your assistance."
)

3. For Realtime Models If you are using a realtime model, you can use the following JavaScript/TypeScript example:

session.conversation.item.create(
  llm.ChatMessage.create({
    role: llm.ChatRole.ASSISTANT,
    text: "Hello, I’m calling to ask you a question about ice cream. Do you prefer chocolate or vanilla?",
  }),
);

session.response.create();

Choose the method that best fits your use case. The

session.say()

method is ideal for fixed messages, while

generate_reply()

is better for dynamic, context-aware responses. Sources: Agent speech and audio | LiveKit Docs | delay-handling/agent.py | tests/test_agent_session.py | Voice AI quickstart | LiveKit Docs | packages/sip--make-a-call/src/agent.ts