Conclave: Research Team Develops Multi-LLM Debate Framework for Enhanced Code Review
Key Takeaways
- ▸Conclave introduces a structured debate mechanism where multiple LLMs deliberate through LEAD, SUPPORT, ALIGN, BUILD, and CHALLENGE rounds before reaching consensus decisions
- ▸The framework intelligently handles varying context windows across different models, auto-truncating debate history to prevent crashes while maintaining information quality
- ▸Cost scales linearly with team size (3 models = 3× expense per message), but free tier options using local models and existing CLI authentication help offset expenses
Summary
Conclave is a new experimental research framework that enables multiple large language models to debate and deliberate before implementing code solutions. Built on top of Anthropic's OpenCode platform, the tool uses a structured debate format with LEAD, SUPPORT, ALIGN, BUILD, and CHALLENGE signals across multiple rounds to reach consensus on solutions. The system works with existing CLI tools like Claude Code, Gemini CLI, and others, requiring no additional API keys beyond existing subscriptions.
The framework demonstrates that collaborative LLM deliberation can catch architectural flaws, edge cases, and security issues that a single model might miss. Users can create multiple named teams with different model combinations, mixing providers from OpenAI, Anthropic, DeepSeek, Google, NVIDIA, Groq, and Ollama. The system intelligently manages context windows by auto-truncating debate threads per model capabilities, allowing high-capacity models to see complete history while smaller models receive signal summaries.
While promising, Conclave is explicitly positioned as experimental research rather than production software. The tradeoff for improved output quality is increased cost and latency: a 3-model team using 3 debate rounds generates 9 API calls per message, multiplying costs proportionally. The project is open to feedback and contributions, with plans for context optimization, live streaming, and autonomous sub-team splitting for complex tasks.
- Works as an open-source layer on top of existing tools, supporting mixed model teams across OpenAI, Anthropic, Google, DeepSeek, and other providers without requiring new API keys
Editorial Opinion
Conclave represents an intriguing shift in how we think about LLM problem-solving—from relying on the best single model to leveraging collaborative reasoning across diverse AI systems. The framework's ability to surface edge cases and security issues through deliberation is compelling, though the cost-quality tradeoff will likely limit adoption to complex tasks where the added expense is justified. As an open research experiment, Conclave could establish valuable patterns for multi-model workflows in production systems.
