Subject: Handling in-flight user speech during multi-agent handoff in AgentSession Hi LiveKit team, I’m implementing multi-agent workflows using AgentSession and handoffs (based on: https://docs.livekit.io/agents/logic/agents-handoffs/). I’m running into an issue with in-flight user speech during agent handoff and wanted to understand the recommended pattern. ### Scenario - Agent1 is actively listening to the user (streaming STT). - Mid-conversation, Agent1 decides to hand off control to Agent2. - During the transition, the user is still speaking (partial utterance, not yet finalized). - When Agent2 takes control, that incomplete / in-flight speech is lost and never reaches Agent2. ### Problem Because the pipeline is streaming (audio → STT → turn detection → agent), it seems the partially processed speech is tied to Agent1’s context and does not get transferred. This leads to: - Lost user input - Broken conversational continuity - Poor UX during transfers ### Question What is the recommended way to ensure that Agent2 receives the in-progress user speech when a handoff occurs? Specifically: 1. Is there a built-in mechanism to buffer or transfer partial transcripts between agents? 2. Should we be implementing a shared transcript/state layer outside the agent (e.g., via context or custom RoomIO)? 3. Is the correct approach to delay handoff until turn completion (via turn detection), or is there a way to safely hand off mid-utterance? 4. Are there hooks in the pipeline (STT stream / transcription events) that we can use to replay or forward partial speech to the next agent? ### Goal We want seamless handoffs where the user can continue speaking naturally without needing to repeat themselves, even if the agent switches mid-sentence. Any guidance or recommended architecture patterns would be greatly appreciated. Thanks!

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