Agent Debate

# Agent Debate Skill

Spawn multiple sub-agents to debate approaches and converge on the best solution.

## Overview

Uses parallel sub-agents with file-based coordination to simulate adversarial debate. Each agent investigates independently, writes findings, then a synthesis agent reviews all positions and picks the winner.

## Pattern: Competing Hypotheses

Best for: architecture decisions, debugging, strategy, trade analysis

### How It Works

1. **Lead defines the question** and spawns 2-4 debate agents
2. Each agent writes their position to `plans/debate-{id}/agent-{n}.md`
3. A synthesis agent reads all positions and produces a verdict
4. Lead reviews and acts on the verdict

### File Structure

```
plans/debate-{topic}/
├── question.md          # The question being debated
├── agent-1.md           # Agent 1's position
├── agent-2.md           # Agent 2's position
├── agent-3.md           # Agent 3's position (optional)
├── rebuttal-1.md        # Agent 1's rebuttal (round 2)
├── rebuttal-2.md        # Agent 2's rebuttal (round 2)
├── synthesis.md         # Final synthesis and verdict
└── decision.md          # Lead's final decision
```

### Usage

#### Single Round (Fast)
3 agents, one round, synthesis. ~5 minutes.

```
1. Write question to plans/debate-{topic}/question.md
2. Spawn 3 agents simultaneously:
   - Agent A: "Argue FOR approach X. Read plans/debate-{topic}/question.md. Write your position with evidence to plans/debate-{topic}/agent-1.md"
   - Agent B: "Argue FOR approach Y. Read plans/debate-{topic}/question.md. Write your position with evidence to plans/debate-{topic}/agent-2.md"
   - Agent C: "Argue FOR approach Z. Read plans/debate-{topic}/question.md. Write your position with evidence to plans/debate-{topic}/agent-3.md"
3. Wait for all to complete
4. Spawn synthesis agent:
   "Read all positions in plans/debate-{topic}/. Score each on: feasibility (1-10), risk (1-10), speed (1-10), quality (1-10). Write verdict to plans/debate-{topic}/synthesis....