Thanks — this is a great use case. I’m responding to your two points: confirming the multi-agent setup and giving the specific event names you can use to detect when one bot finishes speaking so the other can start. Short answer • Yes — you can have 3 participants in a room: two agents (bots) and one human listener. • Use these events to detect when a bot has finished speaking: ◦

agent_state_changed

— watch for

new_state == "idle"

. ◦

SpeechHandle.wait_for_playout()

(or the handle’s done/callback) — signals accurate TTS/playout completion. How to coordinate the two bots (recommended flow) 1. Create two independent AgentSession instances — one per bot. 2. When Bot A begins speaking, you can track that via

agent_state_changed

(

new_state == "speaking"

). 3. When Bot A finishes: ◦ Option A (accurate): await the returned

SpeechHandle

and call

await handle.wait_for_playout()

(or use

handle.add_done_callback

) — then trigger Bot B’s

say()

or

generate_reply()

. ◦ Option B (state-based): listen for

agent_state_changed

with

new_state == "idle"

and then start Bot B. 1. Repeat for handoff back to Bot A. Route each bot’s audio into the room so the human participant can hear both. Which method to use • Use

SpeechHandle.wait_for_playout()

for the most reliable detection of actual audio playback completion. • Use

agent_state_changed

with

idle

for simpler state-driven coordination (less precise if there’s buffering/latency). If you’d like, I can: • Show a concise code example (Python or TypeScript) that wires two AgentSessions together using

SpeechHandle.wait_for_playout()

and routes audio into the room. • Or provide a small state-machine pattern using

agent_state_changed

events. Which would you prefer — a code sample in Python or TypeScript? Sources: Pipeline nodes and hooks | LiveKit Documentation | Turns overview | LiveKit Documentation | How to detect when an agent has finished speaking | LiveKit