Multi-Agent Room

A Multi-Agent Room is a collaborative chat space where three AI agents work together to analyse a topic you bring to them:

Agent	Role
Facilitator	Opens the discussion, structures the problem, asks clarifying questions
Analyst	Provides in-depth analysis, data points, and counterarguments
Summarizer	Synthesises the conversation into conclusions and actionable next steps

You post a topic or question. The agents respond in sequence, each building on the previous agent's output. The conversation is real-time via WebSocket — responses appear as they arrive.

---

How It Works

Under the hood, an Agent Room is a Channel (type='meeting_agents') backed by three configured AgentSkills. This means:

• It uses the same real-time channel infrastructure as the Channels feature
• Each agent's response is stored as a channel message (persistent, searchable)
• The room can be joined by multiple human participants who can read and contribute
• You can configure each agent with different system prompts, knowledge bases, and LLM models

---

Step 1 — Create the Agent Skills

Before setting up an Agent Room plugin, you need three AgentSkills — one for each role.

1. Go to Workspace → Skills → + New Skill.
2. Create a Facilitator skill:
   • System prompt: You are a meeting facilitator. Your job is to structure the discussion, surface key questions, and keep the conversation focused. Be concise. Ask one targeted clarifying question if needed.
   • Attach any knowledge bases relevant to your team's work
3. Create an Analyst skill:
   • System prompt: You are a strategic analyst. Given the facilitator's framing, provide a thorough analysis: key findings, risks, data points, and alternative perspectives. Be specific and factual.
4. Create a Summarizer skill:
   • System prompt: You are an executive summarizer. Given the facilitator's framing and analyst's findings, produce: (1) a 2-sentence summary, (2) 3-5 specific action items with owners. Be concrete and actionable.
5. Copy each skill's ID (shown in the URL: /agent-builder/skills/{id}).

---

Step 2 — Add the Agent Room Plugin

1. Admin Panel → Settings → Meetings → + Add Plugin → agents
2. Fill in:

Field	Value
Name	e.g. Strategy Agent Room
Facilitator Skill ID	UUID from Step 1
Analyst Skill ID	UUID from Step 1
Summarizer Skill ID	UUID from Step 1

3. Click Save.

---

Step 3 — Start a Room

1. Go to Meetings in the sidebar.
2. Click + New Meeting → select your Agent Room plugin.
3. Enter a title (e.g. "Q3 Product Strategy Discussion").
4. Click Start.

You are redirected to the agent room page (/meetings/agents/{room_id}).

---

Using the Agent Room

The room looks like a chat interface:

┌──────────────────────────────────────────┐
│  Strategy Agent Room                     │
│  Facilitator · Analyst · Summarizer      │
├──────────────────────────────────────────┤
│  You: Should we expand into APAC in Q4? │
│                                          │
│  Facilitator: Great question. Before     │
│  diving in, let's clarify: are we        │
│  looking at APAC broadly or specific     │
│  markets like Japan and Singapore?       │
│                                          │
│  Analyst: The APAC expansion opportunity │
│  is significant — the region accounts    │
│  for 45% of global SaaS growth...        │
│                                          │
│  Summarizer: Summary: APAC expansion     │
│  in Q4 is viable with focused entry in  │
│  Singapore. Action items:                │
│  - Alice: Market sizing by Sept 15       │
│  - Bob: Legal entity research by Oct 1  │
├──────────────────────────────────────────┤
│  Post a topic for the agents to discuss… │
└──────────────────────────────────────────┘

Posting a message triggers the full Facilitator → Analyst → Summarizer chain. Each response appears as it completes (real-time streaming via WebSocket).

Multiple humans can join the same room — all see the same conversation. Additional humans can post follow-up questions and the agent chain will re-run.

---

Viewing History

Navigate to Meetings → room → Open Agent Room, or directly to /meetings/agents/{room_id}. The full message history loads on open.

All messages are stored in the backing channel, so they survive page refreshes and server restarts.

---

Generating Meeting Notes

Agent Rooms are text-only — there's no audio recording to feed into the usual Meeting Notes upload flow. Instead, the room page has its own Generate Notes button in the header:

1. Have a conversation with the agent chain (post at least one topic).
2. Pick a model and click Generate Notes.
3. The full conversation (every Facilitator/Analyst/Summarizer exchange) is assembled into a transcript client-side and sent to POST /api/v1/meetings/notes/from-transcript, which summarizes it the same way an uploaded recording would — skipping Whisper entirely since it's already text.
4. You're taken to the resulting note, and it also appears in Meetings → Notes like any other.

This is genuinely useful given the Summarizer agent's whole job is already producing a synthesis — "Generate Notes" turns that into a persisted, searchable record instead of something that only lives in the room's chat history.

---

Tips for Better Results

Provide context-rich topics — the more specific your initial message, the more useful the agent chain output:

# Too vague:
Should we expand?

# Better:
We're considering expanding our B2B SaaS product into the APAC region (specifically
Japan and Singapore) in Q4 2026. Our current ARR is $2M, team size is 12. Should we
prioritize Japan or Singapore first, and what are the key risks?

Use knowledge bases — attach relevant documents (market research, financial models, past meeting notes) to the Analyst skill. The Analyst will cite them in its response.

Customize the agent chain — each skill supports full system prompt customization, knowledge base injection, and LLM model selection. Swap a skill's model to claude-opus-4-8 for deeper analysis on complex topics.

---

Agent Room vs Channels

	Agent Room	Channel
Participants	Humans + 3 AI agents	Humans only (or AI via tool calls)
Message flow	Sequential agent chain per human message	Free-form
Purpose	Structured analysis of a topic	Team communication
Persistence	Yes (channel messages)	Yes
Real-time	Yes (WebSocket)	Yes (WebSocket)

Agent Rooms are built on top of Channels — you can access the raw channel via session.channel_id if you need direct channel API access.

---

Config Schema Reference

Key	Type	Required	Description
facilitator_skill_id	string	Yes	AgentSkill UUID for the Facilitator
analyst_skill_id	string	Yes	AgentSkill UUID for the Analyst
summarizer_skill_id	string	Yes	AgentSkill UUID for the Summarizer

---

Troubleshooting

"Failed to create agent room channel" error → The Channel creation failed, usually due to a permissions issue. Verify that the logged-in user has channel creation permissions in Admin Panel → Settings → Permissions.

Agents don't respond after posting a message → The agent chain runs asynchronously. If there's no response after 30 seconds, check server logs for errors. Common causes: LLM provider not configured, or Ollama/your model provider is not running. A missing or invalid skill ID is caught earlier now — see below.

"Facilitator/Analyst/Summarizer Skill ID does not match any existing AgentSkill" when saving the plugin → The three Skill IDs are validated against real AgentSkill records at Save time, not just checked for presence. This catches a typo'd or since-deleted skill immediately, instead of it silently failing later — previously a bad ID here only surfaced as [Agent error: ...] posted directly into a live room mid-conversation. Copy the correct ID from Workspace → Skills → (skill) → URL.

Wrong skill responding (e.g. Facilitator sounds like the Analyst) → Skill IDs may have been swapped (all three are individually valid, so Save won't catch this). Verify the IDs in Admin Panel → Settings → Meetings against the skill names in Workspace → Skills.

Messages appear in wrong order → Timestamps are based on when each agent's response completes. In a slow LLM environment, responses may arrive out of order. This is cosmetic only — the chain always executes Facilitator → Analyst → Summarizer in sequence.