Moderated Multi-Agent Discussion

Start a structured discussion with AI agents, each with a unique persona.

Pre-flight Check

Before any other action, verify the Agent Teams environment:

1. •

   Check environment variable: The discuss feature requires Agent Teams. If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is not set to 1, inform the user:

   code

   The discuss feature requires Agent Teams to be enabled.
   
   Add this to your .claude/settings.json:
   {
     "env": {
       "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
     }
   }
   

   Then STOP - do not proceed with discussion setup.

Execution

1. •

   If no topic provided: Use AskUserQuestion to interactively gather:

   • •Discussion topic (required)
   • •Number of participants (2-8, default: 4)
   • •Whether it's a debate (pro/con) or open discussion
   • •Any specific roles or perspectives to include
2. •

   Load protocol: Read agents/discuss-lead.md to load the full discussion lead protocol. Execute the protocol directly - do NOT spawn it as a subagent.

3. •

   Analyze topic: Determine team composition (roles, debate mode detection) and prepare persona inputs.

   • •Identify the professional domain and relevant diversity axes for the topic.
   • •Demographics: If geographic origin matters (e.g., global industry practice, policy comparison): estimate practitioner origin distribution, pass demographics: { origin_weights: { ... }, outlier_ratio: 0.2 } to _1_seed. If origin is not the relevant axis: omit demographics and encode diversity directly as a controversy axis.
   • •Briefs: Write a 1-2 sentence background differentiation guide per slot (e.g., "20-year veteran with regulatory background", "Early-career startup founder"). These seed each persona's Expertise section.
   • •Name cultures: Assign a distinct name_culture per agent (no duplicates). Pick from diverse regions (e.g., Korean, Nigerian, Brazilian, German, Indian, Japanese, Egyptian). If demographics provided suggested_origin, use that instead.
4. •

   Generate personas: Spawn persona-generator agents in parallel (one per role, model: "sonnet"). Include from each assignment:

   • •brief (from step 3) as Expertise seed.
   • •name_culture (from step 3 or suggested_origin) - never omit.
   • •positions and tone from _1_seed assignments.
   • •If is_outlier: add context in brief (e.g., "unusual background for this domain - give a compelling career path").
   • •Leave gender, age, and other details to the persona-generator LLM.
5. •

   Initialize: Call discuss_lead({ op: "_2_create", ... }) with generated personas → get session_id

6. •

   Spawn teammates: Create Agent Team coral-dc-{session_id}, spawn discussant teammates

7. •

   Run discussion: Execute discuss_lead(_3_step) rounds - bid collection → winner resolve → speech escalation loop until termination

8. •

   Synthesize: Call discuss_lead({ op: "_7_end", ... }), read full transcript via _4_transcript, present structured summary

Termination Policy

Default: Keep calling _3_step rounds until the system returns phase=ended (automatic termination via epoch exhaustion).

Early manual termination (_7_end) is allowed only when ALL conditions are met:

1. •Every agent has spoken at least twice
2. •Second-round speeches reference or build on earlier positions rather than introducing entirely new topics
3. •At least one agent has explicitly proposed a synthesis or convergence point
4. •The most recent speech does not contain an unanswered question directed at other participants

Before calling _7_end, always check:

• •Is there an agent who hasn't had a second turn yet? If so, run more _3_step rounds instead.
• •Does the last speech pose a question? If so, run more rounds - ending on an unanswered question cuts off dialogue unfairly.

Context Enhancement

From the user's request, identify:

• •Discussion topic (required - gather interactively if not provided)
• •Preferred team size (default: 4–6 agents)
• •Whether it's a pro/con debate (triggers debate mode)
• •Any specific perspectives or roles requested
• •Any controversy hints provided via --hints (pre-specified axes to include in Phase 1 analysis)

Error Policy

If agents/discuss-lead.md cannot be read, report the error to the user.