MAS Harness 🧠

Multi-Agent System Architecture Evolution Engine An autonomous, self-improving multi-agent system that continuously designs, tests, and optimizes agent architectures through closed-loop reinforcement learning.

---

🎯 What is MAS Harness?

MAS Harness is an autonomous AI scientist that runs 24/7 to discover optimal multi-agent system architectures. Unlike static architectures, MAS Harness:

• 🔄 Continuously Evolves: Automatically designs new agent topologies based on benchmark performance
• 📊 Data-Driven: Makes decisions based on objective metrics (success rate, token efficiency, latency)
• 🛡️ Self-Contained: Runs entirely on provided compute resources without human intervention
• 📈 Convergence-Aware: Detects when a paradigm hits diminishing returns and triggers paradigm shifts

---

🏗️ Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    MAS Evolution Engine                      │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐ │
│  │   OODA       │    │  Benchmark    │    │  Resource    │ │
│  │   Loop       │◄──►│  Suite        │◄──►│  Monitor     │ │
│  │              │    │  (16 Tasks)   │    │              │ │
│  └──────────────┘    └──────────────┘    └──────────────┘ │
│         │                   │                   │         │
│         ▼                   ▼                   ▼         │
│  ┌──────────────────────────────────────────────────────┐  │
│  │              Agent Architecture Layer                 │  │
│  │   ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐    │  │
│  │   │Planner │  │ Worker │  │Reviewer│  │Memory  │    │  │
│  │   │ Agent  │  │ Agents │  │ Agent  │  │ Agent  │    │  │
│  │   └────────┘  └────────┘  └────────┘  └────────┘    │  │
│  └──────────────────────────────────────────────────────┘  │
│                          │                                  │
│                          ▼                                  │
│  ┌──────────────────────────────────────────────────────┐  │
│  │              MiniMax M2.7 Model Backend               │  │
│  └──────────────────────────────────────────────────────┘  │
│                                                             │
└─────────────────────────────────────────────────────────────┘

---

📋 Benchmark Tasks

The system evaluates architectures across 5 dimensions:

---

📊 Performance Metrics

---

🚀 Quick Start

Prerequisites

• Python 3.12+
• MiniMax API Key
• 4+ CPU cores, 4GB+ RAM
• GitHub Personal Access Token (for versioning)

Installation

# Clone the repository
git clone https://github.com/xiangbianpangde/mas-harness.git
cd mas-harness

# Install dependencies
pip install psutil

# Configure environment
export MINIMAX_API_KEY="your-api-key"
export GITHUB_TOKEN="your-github-token"

Run Baseline Benchmark

python3 src/mas_v1_single.py

Monitor Evolution

# Check current status
python3 monitor/resource_monitor.py

# View latest results
cat benchmark/results/latest.json | jq '.success_rate, .avg_score'

---

📁 Project Structure

mas-harness/
├── README.md              # This file
├── SOUL.md                # Core directives & constraints
├── AGENTS.md              # Agent workspace conventions
├── HEARTBEAT.md           # Autonomous heartbeat tasks
│
├── EVOLUTION_HISTORY.md   # Architecture changelog
│
├── src/                   # Architecture implementations
│   ├── mas_v1_single.py   # v1.0: Single-agent baseline
│   └── mas_v2_*.py        # v2.0+: Evolved architectures
│
├── benchmark/             # Testing infrastructure
│   ├── mas_benchmark.py   # Benchmark suite (16 tasks)
│   └── results/           # Test results
│
└── monitor/               # Resource monitoring
    └── resource_monitor.py

---

🔄 Evolution Process

OODA Core Loop

┌─────────────────────────────────────────────────────────┐
│                      OODA Loop                          │
├─────────────────────────────────────────────────────────┤
│                                                         │
│   ┌─────────┐                                           │
│   │ OBSERVE │ ←─ Resource Monitor + Benchmark Results   │
│   └────┬────┘                                           │
│        ▼                                                │
│   ┌─────────┐                                           │
│   │ ORIENT  │ ←─ Ablation Analysis + Bottleneck ID     │
│   └────┬────┘                                           │
│        ▼                                                │
│   ┌─────────┐                                           │
│   │  DECIDE │ ←─ Architecture Change or Paradigm Shift │
│   └────┬────┘                                           │
│        ▼                                                │
│   ┌─────────┐                                           │
│   │   ACT   │ ←─ Execute Test + Collect Metrics         │
│   └────┬────┘                                           │
│        │                                                │
│        └──────────────────────────────────────────────►  │
│                      (Loop Continues)                     │
└─────────────────────────────────────────────────────────┘

Convergence Detection

When 10 consecutive generations show <1% improvement:

1. Current best architecture is tagged as a major version (v1.0, v2.0...)
2. Full research report generated
3. New paradigm launched with different topology

---

🛡️ Safety & Constraints

Hard Limits

Red Lines (Never Cross)

• ❌ No network penetration testing
• ❌ No privilege escalation
• ❌ No system directory modification (/etc, /bin, /root)
• ❌ No malicious code execution (crypto miners, DDoS bots)
• ❌ No data exfiltration

---

📈 Version History

---

🔧 Contributing

This is an autonomous system - no human contribution is expected or desired. The repository serves as:

• 📜 Archive of evolution history
• 📊 Benchmark for architecture evaluation
• 📖 Documentation of discovered architectures

For questions or issues, please refer to the archived research papers generated with each major release.

---

📜 License

MIT License - See LICENSE for details.

---

🧭 Navigation

• SOUL.md - Core directives the system follows
• EVOLUTION_HISTORY.md - Detailed changelog
• benchmark/mas_benchmark.py - Task definitions
• monitor/resource_monitor.py - Resource tracking

---

Built with autonomous evolution in mind. No humans were harmed in the design of this system. 🤖