Council - Multi-Agent Debate

Orchestrates parallel debates across multiple AI agents (codex, gemini, super-agent) with a Chair LLM synthesizing responses across rounds.

When to Use

• "Get multiple perspectives on this"
• "Have the agents debate this approach"
• "Review this plan with the council"
• "What do different models think about X?"

How It Works

┌─────────────────────────────────────────────────────────────┐
│                      ROUND N                                │
├─────────────────────────────────────────────────────────────┤
│  1. All agents receive the question + previous context      │
│  2. Agents respond IN PARALLEL (O(N) not O(N²))             │
│  3. Chair synthesizes responses                             │
│  4. Chair decides: CONTINUE or DONE                         │
│  5. If CONTINUE, agents get synthesis + other views         │
└─────────────────────────────────────────────────────────────┘

Each agent maintains its own session via --resume, enabling true multi-round debate where agents remember what they said.

Usage

# Basic question
~/.claude/skills/council/scripts/council "What's the best approach for X?"

# Specify agents (default: gemini,super)
~/.claude/skills/council/scripts/council --agents codex,gemini,super "Review this"

# More rounds for complex debates
~/.claude/skills/council/scripts/council --max-rounds 5 "Debate the tradeoffs"

# Verbose output (see agent responses in real-time)
~/.claude/skills/council/scripts/council --verbose "Analyze this architecture"

# From file
PROMPT_FILE=question.md ~/.claude/skills/council/scripts/council

Options

Flag / Variable	Description
--agents, -a	Comma-separated agents (default: gemini,super)
--max-rounds, -r	Maximum debate rounds (default: 3)
--verbose, -v	Show agent responses in real-time
PROMPT_FILE	Read question from file
COUNCIL_AGENTS	Default agents via environment

Available Agents

Agent	Strengths	Notes
gemini	Fast, good UX/behavior reasoning	Enabled by default
super	Architectural, can use tools	Enabled by default
codex	Technical correctness, code review	Often rate-limited

Output

• Progress/status goes to stderr (council convened, rounds, decisions)
• Final synthesis goes to stdout (pipeable)

# Capture just the final answer
answer=$(~/.claude/skills/council/scripts/council "Question" 2>/dev/null)

# See the debate unfold
~/.claude/skills/council/scripts/council --verbose "Question"

Extending

Add new agents by updating AGENT_SCRIPTS in the script:

AGENT_SCRIPTS = {
    "codex": Path.home() / ".claude/skills/codex-agent/scripts/codex-agent",
    "gemini": Path.home() / ".claude/skills/gemini-agent/scripts/gemini-agent",
    "super": Path.home() / ".claude/skills/super-agent/scripts/super-agent",
    "cursor": Path.home() / ".claude/skills/cursor-agent/scripts/cursor-agent",  # Add new
}

Agents must support:

• Prompt as final argument
• --resume <session_id> for continuity
• Output [session_id: ...] for session tracking