Multi-Agent Adversarial Debate System
Beyond single-model AI — structured argumentation for balanced, well-reasoned decisions
Inspiration • Features • Quick Start
A multi-model debate interface inspired by Andrej Karpathy's LLM Council.
Instead of querying a single LLM, this system routes one prompt to multiple models, lets them review each other's responses, and produces a final answer through structured synthesis.
The original LLM Council by Andrej Karpathy is a lightweight experiment for:
- Comparing multiple LLM outputs
- Peer-review between models
- Consensus-based answers
This project builds on that idea and adds structure and extensibility.
- Explicit roles: Pro, Con, Judge
- Optional multi-round debate instead of single pass
- Memory support (PostgreSQL + FAISS) for past debates
- Uncertainty & confidence signals in final answers
- LLM routing with OpenRouter and local Ollama fallback
- Basic governance hooks (audit, rollback-ready design)
User Query
→ Pro Agent
→ Con Agent
→ Judge Agent
→ Final Response
- Python 3.10+
- Node.js 18+ (UI)
- OpenRouter API key (optional if using Ollama)
git clone https://github.com/yourusername/llm-council.git
cd llm-council
pip install -r requirements.txt
python -m src.api.main- Local-first, not a hosted service
- Models are interchangeable
- Cost depends on selected LLMs
- Intended for experimentation and system design exploration
MIT License — See LICENSE