Thanks — this is a great summary of the problem and your current workaround. I’ll respond to each of your numbered questions and then give a concise recommended pattern you can implement instead of polling. Acknowledgement: you’re asking how to detect when the agent schedules the next speech after your tool returns (so you can await the new turn and then delete the room), and how to avoid polling while still supporting the “play goodbye → if interrupted → let agent respond → then hang up” flow. Answers 1) Is there an event or awaitable to know when the next speech handle has been scheduled after a tool returns? • Yes. AgentSession emits a speech_created event whenever a new speech is initiated. You can attach a one‑time listener (for example

session.once("speech_created", callback)

) to be notified when the next speech is created instead of polling

session.current_speech

. 2) What is the correct pattern for "play goodbye → if interrupted → let agent respond → then hang up"? • Return from the tool immediately after initiating the goodbye speech. If the goodbye is interrupted, register a one‑time

speech_created

listener. When

speech_created

fires, await the new speech handle’s playout completion using

candidate.wait_for_playout()

and then run your teardown (e.g.,

delete_room()

). This ensures the SDK can schedule the next response (S2) after the tool returns and you wait for the agent turn to finish before hanging up. 3) Is there a way to await the next agent turn completing, rather than watching current_speech? • There isn’t a single “wait_for_next_turn” helper, but the pattern above (listen for

speech_created

→ get the new speech handle → await its

wait_for_playout()

) is the correct way to await the agent’s next turn without polling. 4) Does AgentSession expose any wait_for_next_speech() style API? • No. AgentSession does not expose a

wait_for_next_speech()

API. Use the

speech_created

event and then await the returned speech handle’s

wait_for_playout()

. Recommended implementation sketch 1. When you play the goodbye speech from inside the tool, return from the tool immediately (do not block waiting for S2). 2. If the goodbye speech was interrupted: ◦ Use

session.once("speech_created", callback)

to catch the next speech. ◦ In the callback, obtain the new speech handle (the callback should receive it or you can read

session.current_speech

) and await

candidate.wait_for_playout()

with a reasonable timeout. ◦ After awaiting the new speech playout completion, run your teardown (delete room etc.). 1. If goodbye was not interrupted, carry out teardown immediately. Notes and timing tips • Give a slightly larger window for the next speech to appear (your 1s polling window was tight). The

speech_created

event removes the need to poll and avoids timing races. • Use a timeout on

wait_for_playout()

(e.g., 5–10s) to avoid hanging if the agent’s speech never appears or stalls. If you’d like, I can: • Show a short async code snippet using

session.once("speech_created", ...)

and

candidate.wait_for_playout()

. • Help adapt this pattern into your existing

hangup()

/

_wait_for_reply_and_teardown()

flow. Which would you prefer? Sources: Agent speech and audio | LiveKit Documentation | Agent session | LiveKit Documentation | Agents telephony integration | LiveKit Documentation | How to detect when an agent has finished speaking | LiveKit