Multi-agent turn coordination — text-stream ack protocol vs native SDK approach?

Agents
1

Hello,

We have two agents in the same LiveKit room using livekit-agents~=1.5. Agent A speaks first, then sends a text payload to Agent B via send_text(topic=“finding”). Agent B reacts via session.say() + wait_for_playout(), then sends send_text(“done”, topic=“turn-ack”). Agent A awaits the ack before proceeding to the next item.

This works well in testing (zero overlap between agents). We chose this over VAD-based gating because TTS has natural inter-sentence pauses that triggered false “end of speech” signals.

Questions:

  1. Is there a native SDK mechanism for explicit agent-to-agent turn coordination we should be using instead? (e.g., something in the handoff/task system, or a built-in signaling pattern?)
  2. Any concerns with using send_text() as a low-frequency signaling channel? (Our rate is ~1 message per 15-20s.)
  3. For scaling to 3+ agents in the same room, would you recommend a different pattern?

Thanks!

2

There isn’t a lower-level “turn lock” primitive in the SDK specifically for agent-to-agent coordination. The native pattern for structured multi-agent control is agent sessions with handoffs, tasks, and task groups, where one controlling agent transfers execution explicitly rather than relying on media timing. See Workflows and the linked Agents & handoffs section.

Using send_text() as a low-frequency signaling channel (1 message per 15–20s) is fully reasonable. It rides on the same reliable data mechanisms as other text streams and is appropriate for explicit coordination.

For 3+ agents, consider a single controlling agent (or task group) orchestrating handoffs instead of peer-to-peer acks, which keeps flow centralized and testable.

Are these agents meant to act as distinct “personas” in one conversation, or as cooperative background workers?

3

Hi,

Thanks for the tips and doc references, this is really helpful. I think we now need to switch to this approach.

Currently we have one persona with multiple agents under the hood, same voice. STT+VAD is needed since a real human can join.

Two follow-up questions: for a future use case we want 2–4 independent personas in one room. I think we can drive interruptions programmatically via @function_tool rather than STT and VAD. Do you think that’s a good idea or is there a better approach?

Thanks

4

I am not sure. It really comes down to your use case and what you are really trying to achieve. How those “personas” interact, or if there only one active at a time, etc.