You're asking how to stop bots from talking to themselves and improve script adherence — here's a concise list of actions you can take now. 1. Prevent agents responding to agent messages • In your conversation hook, ignore messages from agent participants. Example (Python):

async def on_user_turn_completed(self, turn_ctx, new_message):
    if new_message.participant_kind == rtc.ParticipantKind.PARTICIPANT_KIND_AGENT:
        raise StopResponse()

1. Restrict which participant kinds the agent listens to • Start sessions so agents only accept

STANDARD

(and

SIP

if needed) participants:

await session.start(
  room_input_options=RoomInputOptions(
    participant_kinds=[rtc.ParticipantKind.PARTICIPANT_KIND_STANDARD]
  )
)

1. Use tools for internal logic instead of spoken turns • Run internal checks or follow-ups as silent function tools so they don’t generate speech:

@function_tool()
async def internal_check(self):
    result = await self.run_internal_logic()
    return result, "Internal result"

1. Only call speech APIs when user input requires it • Avoid generating replies from internal events; use

session.say()

or

generate_reply()

only for deliberate, user-facing responses. Optionally disable interruptions:

session.say("Hello", allow_interruptions=False)

1. Validate message origin and identity • Check `participant_identity`/metadata on incoming messages; reject or ignore messages that originate from the bot itself. If you want, tell me which SDK/language you’re using (Python or Node), and I’ll give the exact code snippets to drop into your agent. Sources: Pipeline nodes and hooks | LiveKit docs | Prompting guide | LiveKit docs | How to test your agent using another agent | Agents and handoffs | LiveKit docs | Agent speech and audio | LiveKit docs