Once my realtime agent completes a conversation, it automatically loops to the beginning without user input. How do I prevent this?

Thanks for reaching out, Matthew. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

This is happening without the user speaking at all. So audio input is not the issue, nor is turn detection.

It’s great that audio input and turn detection aren’t causing the issue! To help you stop the agent from looping, I suggest using manual turn control. This way, you can turn off automatic triggers and have full control over when the agent processes input. Here’s a simple way to set it up: • Set

turn_detection="manual"

in the

AgentSession

constructor to turn off automatic turn detection. • Take charge of the agent’s actions by using methods like

session.commit_user_turn()

to start responses when you want, and

session.clear_user_turn()

to clear any input you don’t want processed. • Double-check that there aren’t any conflicting settings, like server-side turn detection, enabled in your LLM configuration. This approach makes sure the agent only responds when you say so, preventing any unwanted restarts or loops. If you’d like, I’m here to walk you through the setup or answer any other questions! Sources: Turn detection and interruptions | LiveKit Docs | livekit-agents/livekit/agents/voice/agent_activity.py

Is it not the default for an agent that initiates the conversation to not restart the conversation afterwards? I am using an rpc call tool that signals completion to the front-end that fires off midway through the agent's final speech so I cannot cut the session with that tool either.

I understand how important it is to have your agent behave just the way you expect, so let me walk you through this gently. By default, an agent that starts a conversation won’t automatically restart once it’s done. But if you’re using an RPC call tool that signals completion while the agent is still speaking, it might cause some unexpected behavior. Here are a few friendly steps to help you manage this: • Explicitly End the Session: Try using a function tool to close the session right after the agent finishes speaking. For instance,

ctx.wait_for_playout()

can help make sure the agent completes its speech before the session ends. • Suppress Auto-Generated Replies: If your tool triggers a reply you don’t want, you can return

None

from the tool or use

ev.cancel_tool_reply()

within the

function_tools_executed

event to stop any unintended responses. • Manual Turn Control: Consider disabling audio input by default and instead use RPC methods like

start_turn

and

end_turn

to clearly control when the agent listens and responds. Following these steps should help keep the agent from restarting or continuing after it’s done with its task. If anything feels unclear or you’d like a hand setting this up, please don’t hesitate to reach out—I’m here to help! Sources: livekit-plugins/livekit-plugins-aws/livekit/plugins/aws/experimental/realti... | Agent speech and audio | LiveKit Docs | Agent speech and audio | LiveKit Docs | packages/basics/src/uninterruptable.ts | Turn detection and interruptions | LiveKit Docs

1. Explicit ends cut the speech. Not an option 2. Where would I invoke this? I just need it to not loop after the final message which triggers my rpc tool. 3. Absolutely not

I completely understand how important it is to prevent the agent from looping after delivering its final message, while still allowing it to finish speaking naturally. Here are a few friendly approaches you might find helpful: • Coordinate Final Speech: You can use

ctx.wait_for_playout()

or the

SpeechHandle

returned by

say()

. This way, the agent completes its final message smoothly before any other actions come into play. • Use State Management: Try keeping track of a

conversation_ended

flag within your agent’s logic. After the final message is delivered, you can gently stop further replies by raising

StopResponse()

in the

on_user_turn_completed

lifecycle hook. • Leverage on_exit(): Another great option is to implement the

on_exit()

method. This lets you send a one-time farewell message right when the session ends, so there’s no risk of looping. These strategies should help your agent wrap up conversations smoothly and avoid any unintended restarts. If you'd like, I’m here to guide you through implementing any of these solutions—just let me know! Sources: Agents telephony integration | LiveKit Docs | Pipeline nodes and hooks | LiveKit Docs | Agent speech and audio | LiveKit Docs | examples/voice_agents/session_close_callback.py | pipeline-llm/interrupt_user.py