Agent Debate

What This Skill Does

The Agent Debate skill is a powerful orchestration framework designed for the OpenClaw platform that enables complex decision-making through multi-agent adversarial synthesis. Instead of relying on a single AI response, this skill allows the user to spawn multiple sub-agents that investigate, argue, and iterate on a specific problem or architectural question. By leveraging a structured file-based coordination pattern, the system forces agents to document their positions, read their peers' findings, and engage in constructive rebuttal. The process culminates in a formal synthesis report, ensuring that the final output is pressure-tested against alternative viewpoints. This approach significantly reduces the impact of LLM hallucinations and confirmation bias in high-stakes reasoning tasks.

Installation

To integrate this skill into your OpenClaw environment, execute the following command in your terminal: clawhub install openclaw/skills/skills/0xrichyrich/agent-debate

Use Cases

• Architectural Decision Making: Comparing different database schemas or infrastructure designs to identify bottlenecks before implementation.
• Complex Debugging: Having multiple agents analyze a stack trace from different perspectives to isolate the root cause.
• Strategy & Trade Analysis: Evaluating market conditions or product roadmaps by forcing agents to represent opposing strategic interests.
• Security Hardening (Red Teaming): Using the adversarial pattern to build a secure system, where one agent drafts the implementation and the other actively attempts to break it.

Example Prompts

1. "Execute Agent Debate on topic 'DatabaseMigration': Create 3 agents to compare PostgreSQL vs MongoDB for our high-throughput analytical dashboard. Write positions to plans/debate-DatabaseMigration/."
2. "Run Red Team debate for the new authentication module: Agent A writes the logic in proposal.md, then Agent B attempts to find privilege escalation vulnerabilities."
3. "Summarize the consensus for 'UI-Framework-Selection': Read all rebuttal files in plans/debate-UI-Framework-Selection/ and generate a final recommendation based on developer velocity and runtime performance."

Tips & Limitations

• Context Window: Since agents read files to coordinate, ensure your intermediate files remain concise; otherwise, you may hit token limits in the synthesis phase.
• File Management: Always clean up the plans/debate-{topic}/ directories after the final decision is reached to prevent clutter and potential conflict in future runs.
• Cost: Running multiple agents simultaneously consumes more model tokens than a single-agent workflow; use the Single Round approach for quick tasks and save the Two Round/Red Team patterns for mission-critical decisions.