Agent Debate is a multi-agent framework that improves response accuracy and reduces hallucinations by having multiple AI agents independently answer, critique, and refine solutions through structured disagreement.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install multi-agent-debate
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install multi-agent-debate using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Agent Debate?
Agent Debate implements a powerful multi-agent paradigm designed to achieve truth through structured disagreement. By leveraging the 'Wisdom of Crowds,' this skill ensures that multiple agents independently address a prompt before engaging in a multi-round critique process. This peer-review mechanism is specifically engineered to surface contradictions, identify logical fallacies, and catch the 'confident-sounding' wrong answers that often plague single-agent interactions.
Built on top of Openclaw Skills, this framework transforms a simple query into a high-fidelity panel discussion. It is particularly effective for high-stakes decision-making where precision is paramount. By forcing agents to survive the scrutiny of their peers, the final output gains a layer of verification and confidence that single-pass LLM responses cannot match.
Agent Debate Use Cases
The Agent Debate skill is essential for scenarios where accuracy outweighs speed:
- Fact-checking high-stakes claims to ensure data integrity.
- Solving complex mathematical or logic puzzles through independent verification.
- Performing trade-off analysis for architectural or business decisions.
- Identifying and mitigating AI hallucinations in technical documentation.
- Reaching consensus on multi-faceted questions with no single objective answer.
How Agent Debate Works
The Agent Debate process follows a rigorous four-step lifecycle to ensure maximum accuracy:
- Independent Generation: Multiple agents (ideally an odd number) generate answers simultaneously without seeing each other's responses to prevent bias.
- Structured Critique: Each agent reviews the others' work, identifying omissions, weak reasoning, or factual errors, while also performing a self-critique.
- Answer Refinement: Based on the peer feedback, agents revise their initial responses, incorporating valid criticisms or defending their positions with additional evidence.
- Consensus or Synthesis: A final verdict is reached through majority voting or a meta-agent synthesis that combines the strongest elements of all refined answers.
Agent Debate Setup
To integrate Agent Debate into your workflow, ensure your Openclaw Skills environment is active. You can initiate a debate-style reasoning task using the following structure in your agent configuration:
# Define the debate parameters and agent count
openclaw-cli skill add agent-debate
# Configure a 3-agent debate for complex reasoning
openclaw-cli run debate --agents 3 --topic "Analysis of 2008 Financial Crisis"
Agent Debate Data Schema & Taxonomy
Agent Debate organizes its workflow data using a structured taxonomy to track the evolution of the truth:
| Phase | Data Generated | Metadata |
|---|---|---|
| Round 1 | independent_responses[] |
Agent ID, Timestamp |
| Round 2 | peer_critiques{} |
Target ID, Error Tags, Scores |
| Round 3 | refined_responses[] |
Revision Delta, Defense Notes |
| Final | final_verdict |
Confidence Level (High/Med/Low) |
All interactions are logged in Markdown format to maintain a clear audit trail of how consensus was reached.
Agent Debate Advanced Features
Power users can extend the Agent Debate skill within the Openclaw Skills ecosystem using these advanced features:
- Meta-Synthesis: An automated meta-agent that extracts consensus points and identifies remaining points of contention.
- Custom Debate Formats: specialized workflows for Binary Fact-Checks, Step-by-Step Reasoning, or Trade-off Analysis.
- Integration Hooks: Combine with 'Plan First' to debate strategies before execution or 'Team Code' for architectural reviews.
- Confidence Scoring: Automated assessment of consensus levels to alert users when agents remain in significant disagreement.
SKILL.md
Loading
Related Openclaw Skills
Aliyun SLS Openclaw Integration
Provision Alibaba Cloud SLS observability for AI agents with automated collector setup and dashboard creation.
TTS Synthesizer
A versatile Text-to-Speech utility for AI agents to generate natural-sounding voice responses using Microsoft Neural TTS or OpenAI compatible APIs.
Aliyun DevOps Manage
A comprehensive automation skill for managing Alibaba Cloud DevOps (Yunxiao 2020) resources through secure OpenAPI and SDK workflows.
Multi-Cloud Product Docs & API Benchmark
An automated tool for benchmarking and comparing documentation and API quality across seven major cloud service providers.
Weekly Report Generator
An intelligent agent skill that automatically analyzes git commit history to generate and distribute professional weekly status reports.
Alibaba Cloud SLS Log Query
A powerful AI agent skill for querying, analyzing, and troubleshooting logs within Alibaba Cloud Log Service (SLS) using the Python SDK.
Featured*
