🤝 Aegean Consensus

A Byzantine Fault-Tolerant Consensus Protocol for Multi-Agent LLM Systems

Reaching Agreement Among Reasoning LLM Agents with Formal Guarantees

📖 Documentation • 🚀 Quick Start • 🎯 Features • 🏗️ Architecture

---

🌟 Overview

Aegean is a production-ready implementation of the consensus protocol described in the paper "Reaching Agreement Among Reasoning LLM Agents" (arXiv:2512.20184). It enables multiple LLM agents to reach reliable consensus on complex reasoning tasks with formal guarantees of correctness.

Why Aegean?

Traditional multi-agent systems suffer from three critical problems:

Problem	Traditional Approach	Aegean Solution	Performance Gain
🔄 Fixed Rounds	Predetermined iteration limit	Adaptive termination via stability horizon	1.2-20× faster
⏱️ Barrier Sync	Wait for slowest agent	Early termination mechanism	Latency decoupled
⚠️ Temporary Agreement	Output first consensus	Stability guarantee (β rounds)	Robust consensus

Key Results from Paper

• Latency Reduction: 1.2× to 20.2× faster (AIME dataset)
• Token Efficiency: 1.1× to 4.4× fewer tokens
• Accuracy Improvement: +2.3% to +5.1% across benchmarks
• Formal Guarantees: Refinement validity, monotonicity, and termination

---

🎯 Features

Core Protocol

• ✅ Quorum Detection - Byzantine fault-tolerant voting (Section 5.1)
• ✅ Stability Horizon - Prevents premature consensus (Section 5.2)
• ✅ Early Termination - Cancels slow agents after quorum (Section 6)
• ✅ Refinement Rounds - Iterative solution improvement
• ✅ Formal Guarantees - Proven safety and liveness properties

Multi-Framework Support

• 🔧 AutoGen Integration - Native support for Microsoft AutoGen
• 🦅 OpenClaw Integration - Dynamic node activation via message triggers
• 🔌 Custom Agents - Easy-to-implement agent interface
• 🌐 Distributed Architecture - gRPC-based multi-node deployment

Production Ready

• 🚀 FastAPI Service - RESTful API with async support
• 📊 Prometheus Metrics - Built-in monitoring and alerting
• 🐳 Docker/K8s - Container-ready deployment
• 📝 Structured Logging - Debug-friendly with correlation IDs

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────┐
│                    Client Layer                          │
│         Web UI | CLI | Python SDK | REST API            │
└────────────────────────┬────────────────────────────────┘
                         │
┌────────────────────────▼────────────────────────────────┐
│              Aegean Consensus Service                    │
│  ┌────────────────────────────────────────────────────┐ │
│  │  Consensus Coordinator                             │ │
│  │  • Leader Election                                 │ │
│  │  • Quorum Detection                                │ │
│  │  • Stability Tracking                              │ │
│  │  • Early Termination                               │ │
│  └────────────────────────────────────────────────────┘ │
│  ┌────────────────────────────────────────────────────┐ │
│  │  OpenClaw Gateway                                  │ │
│  │  • Dynamic Node Activation                         │ │
│  │  • Lifecycle Management                            │ │
│  └────────────────────────────────────────────────────┘ │
└────────────────────────┬────────────────────────────────┘
                         │
        ┌────────────────┴────────────┬──────────────┐
        │                             │              │
┌───────▼────────┐         ┌─────────▼──────┐  ┌───▼──────┐
│ AutoGen Agents │         │ OpenClaw Nodes │  │  Custom  │
│ (Static Pool)  │         │ (On-Demand)    │  │  Agents  │
└────────────────┘         └────────────────┘  └──────────┘

Design Highlights

• Passive OpenClaw Activation: Nodes start only when needed, saving GPU resources
• Dynamic Registration: Agents register at runtime, not pre-configured
• Direct Communication: HTTP/gRPC for low-latency consensus
• Hybrid Agent Pool: Mix static (AutoGen) and dynamic (OpenClaw) agents

---

🚀 Quick Start

Prerequisites

• Python 3.9+
• Docker (optional, for containerized deployment)
• GPU (optional, for OpenClaw nodes)

Installation

# Clone the repository
git clone https://github.com/your-org/aegean-consensus.git
cd aegean-consensus

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Configure environment
cp .env.example .env
# Edit .env with your API keys

Basic Usage

from aegean import ConsensusCoordinator, AgentRegistry

# Initialize
registry = AgentRegistry()
coordinator = ConsensusCoordinator(
    agent_registry=registry,
    quorum_size=2,
    stability_horizon=2
)

# Run consensus
result = await coordinator.run_consensus(
    consensus_id="task-001",
    task="What is the capital of France?",
    enable_openclaw=True
)

print(f"Consensus: {result.final_solution.answer}")
print(f"Rounds: {result.current_round}")
print(f"Agents: {result.participating_agents}")

Run as Service

# Start the FastAPI service
python main.py

# Or use Docker
docker-compose up -d

# Test the API
curl -X POST http://localhost:8000/api/consensus/solve \
  -H "Content-Type: application/json" \
  -d '{
    "task": "Solve: 2x + 5 = 13",
    "config": {
      "quorum_size": 2,
      "stability_horizon": 2,
      "enable_openclaw": true
    }
  }'

---

📖 Documentation

• Technical Specification - Complete system design and implementation details
• API Reference - REST API documentation
• Deployment Guide - Production deployment instructions
• Contributing - How to contribute to the project

---

🔬 How It Works

The Aegean Protocol (Simplified)

# Algorithm 1 from the paper
def aegean_consensus(task, agents):
    # 1. Leader Election
    leader = elect_leader(agents)
    
    # 2. Collect Initial Solutions
    solutions = collect_solutions(agents, task)
    
    # 3. Refinement Loop
    for round in range(max_rounds):
        # Check quorum (≥ ⌈N/2⌉ agents agree)
        candidate = check_quorum(solutions)
        
        if candidate:
            # Update stability tracker
            if stability_tracker.update(candidate):
                # Stable for β rounds → consensus!
                return candidate
        
        # Refine solutions based on previous round
        solutions = refine_solutions(agents, solutions)
    
    return None  # No consensus reached

Key Mechanisms

1. Quorum Detection (Section 5.1)

Quorum Size α = ⌈N/2⌉ (simple majority)
Example: 5 agents → need 3 to agree

2. Stability Horizon (Section 5.2)

Candidate must maintain quorum for β consecutive rounds
Example: β=2 means 2 rounds of stable agreement

3. Early Termination (Section 6)

Cancel slow agents after collecting α responses
Latency = max(fastest α agents), not max(all agents)

---

🧪 Benchmarks

Performance comparison on standard datasets (from paper):

Dataset	Metric	Baseline	Aegean	Improvement
GSM8K	Latency	41.2s	10.0s	4.1× faster
	Tokens	1,200	1,090	1.1× fewer
	Accuracy	87.5%	89.8%	+2.3%
MMLU	Latency	88.4s	10.0s	8.8× faster
	Tokens	2,700	1,000	2.7× fewer
	Accuracy	76.2%	78.0%	+1.8%
AIME	Latency	202.0s	10.0s	20.2× faster
	Tokens	4,400	1,000	4.4× fewer
	Accuracy	23.3%	28.4%	+5.1%

Note: Results from paper experiments with N=5 agents, α=3, β=2

---

🛠️ Configuration

Basic Configuration

# config/production.yaml
consensus:
  quorum_size: 2          # ⌈N/2⌉ for N agents
  stability_horizon: 2    # β rounds for stability
  max_rounds: 5           # Maximum refinement rounds
  timeout: 300            # Seconds

agents:
  autogen:
    - id: agent_0
      model: gpt-4
      temperature: 0.7
    - id: agent_1
      model: claude-3-opus
      temperature: 0.7

openclaw:
  enabled: true
  cluster_url: http://openclaw-cluster:9000
  timeout: 300

OpenClaw Integration

openclaw:
  enabled: true
  cluster_url: http://openclaw-cluster:9000
  callback_url: http://aegean-service:8000
  activation_mode: passive    # Start nodes on-demand
  registration_mode: dynamic  # Register at runtime
  communication: direct       # HTTP/gRPC (not message queue)

---

🐳 Deployment

Docker Compose

# Start all services
docker-compose up -d

# Services:
# - aegean-service:8000 (Consensus API)
# - openclaw-gateway:9000 (OpenClaw Gateway)
# - prometheus:9090 (Metrics)
# - grafana:3000 (Dashboards)

Kubernetes

# Deploy to K8s cluster
kubectl apply -f k8s/aegean-deployment.yaml
kubectl apply -f k8s/openclaw-deployment.yaml

# Check status
kubectl get pods -l app=aegean

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Development Setup

# Install dev dependencies
pip install -r requirements-dev.txt

# Run tests
pytest tests/

# Run linter
flake8 src/
black src/

# Run type checker
mypy src/

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

📚 Citation

If you use Aegean in your research, please cite the original paper:

@article{aegean2024,
  title={Reaching Agreement Among Reasoning LLM Agents},
  author={[Authors from paper]},
  journal={arXiv preprint arXiv:2512.20184},
  year={2024}
}

---

🙏 Acknowledgments

• Original paper: Reaching Agreement Among Reasoning LLM Agents
• Microsoft AutoGen - Multi-agent framework
• FastAPI - Modern web framework
• OpenClaw - LLM inference platform

---

📞 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Email: aegean-dev@your-org.com

---

Built with ❤️ for the Multi-Agent AI Community

⭐ Star us on GitHub if you find this project useful!