Meetings Guide

Multi-agent meetings — the guide

The meetings card in your instructions carries the rules (when a meeting is right, who you can meet, the start_meeting rule, the do-nots). This guide is the reference behind them.

How meetings work

1. Call start_meeting(topic, agents) — you become the moderator
2. Use direct_to(agents=[...]) to address specific agents — they respond next (in parallel if multiple)
3. If you don't call direct_to, your response broadcasts to all participants
4. All responses are visible to everyone in the transcript, regardless of routing
5. As moderator, call end_meeting to conclude — your response becomes the summary
6. Any agent can call propose_conclude to suggest ending — you (moderator) decide

Turns are strictly sequential. The platform routes your message only when your turn ENDS: the agents you addressed start speaking after your response is complete, and what they say reaches you as transcript in your next turn. So a turn is always the same shape — gather what you need with tools, write your message as response text, call the routing tool once, stop. Nothing you do after the routing call can hear a reply; the platform denies later tool calls in that turn (memory writes are the one exception, so you can still save what the meeting taught you on the way out).

Tool guide

Best practices

• Starting a meeting — CRITICAL: When you call start_meeting, the meeting session starts AFTER your current response completes. You will then receive a separate, dedicated prompt as the moderator to open the discussion inside the meeting. Therefore: your response that calls start_meeting must ONLY contain a brief acknowledgment + the tool call. Do NOT call direct_to, do NOT discuss the topic, do NOT address other agents, do NOT share opinions. Any text or tool calls after start_meeting in the same response happen OUTSIDE the meeting and are wasted — the other agents will never see them. Correct: "Sure, let me set up that meeting." → start_meeting(...).
• As moderator: Open with a clear agenda. Use direct_to to address specific agents for their input, then stop — their answers arrive in your next turn. When done, write the summary, call end_meeting, stop.
• As participant: Be concise (1-3 paragraphs). Address other agents by name. Disagree constructively. Call propose_conclude when you have nothing more to add.
• Do NOT use the Agent tool, background subagents, or delegate during meetings.
• Do NOT respond with just acknowledgments — if you have nothing to add, call propose_conclude.

Example flow

Turn 1 — Moderator: "Let's check store and systems health."
  → direct_to(["home-assistant", "system-admin"])   → turn ends

Turn 2 — Home Assistant + System Admin respond IN PARALLEL with their reports
  → both direct_to(["personal-assistant"])  (report back)   → turns end

Turn 3 — Moderator (both reports now in the transcript): "Here's the summary. [action items]"
  → end_meeting   → turn ends, meeting concluded