🧠 Multi-Agent Coder (From Scratch)

A minimal, fully transparent agentic AI system built in pure Python — no frameworks, no abstractions.

This project demonstrates how multiple specialized agents can plan, write, review, and refine code collaboratively using simple tools and a shared workspace.

Designed for learning, teaching, and experimentation with real agentic patterns.

---

🚀 Features

🧩 Architect Agent

• Breaks down user requests into:
  • File structure
  • Modules
  • Step-by-step engineering plan

🛠️ Engineer Agent

• Writes and updates code
• Saves files to disk
• Executes Python scripts
• Fixes errors based on feedback

🔍 Critic Agent

• Reviews generated code
• Checks correctness
• Suggests improvements
• Ensures consistency

---

🔄 Agent Workflow

Architect → Engineer → Critic → Engineer → Critic → … → Done

This loop is the core pattern behind modern agentic systems.

---

📁 Project Structure

multi-agent-coder/
│
├── agents/
│   ├── base_agent.py      # Core agent class
│   ├── architect.py       # Planning agent
│   ├── engineer.py        # Code-writing agent
│   └── critic.py          # Review agent
│
├── workspace/             # Generated code output
│
├── tools.py               # File I/O + Python execution tools
├── controller.py          # Orchestration loop
└── README.md

⚙️ How It Works

Each agent is a lightweight wrapper around an LLM with:

• A system prompt
• Message history
• A send() method for communication

🧰 Built-in Tools

• 📄 File read/write
• ▶️ Python execution

Everything is intentionally simple and transparent, making it easy to understand and extend.

📜 License

This project is licensed under the MIT License.

⭐ Contributing

If you want to experiment or extend this system:

• Fork the repo
• Add your agent or tool
• Open a PR