This framework enables the dynamic creation of self-improving AI agents capable of editing their own prompts and Python code. It targets developers and researchers seeking to build adaptive AI systems that learn and reuse successful task execution strategies across sessions. The core benefit is the automated generation and validation of functional code snippets for task completion.

How It Works

Microagents are small, dynamically generated agents that respond to user tasks. They are assessed for functionality, and successful agents are stored for reuse, facilitating cross-session learning. This approach leverages large language models (like GPT-4 Turbo) to generate Python code for specific tasks, with a validation phase ("Judge") ensuring only functional agents are cataloged. Agents are persisted in an SQLite database.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Set environment variables for OpenAI API key (or Azure OpenAI credentials).
• Run demo: python main.py
• Run interactive chat: python app.py
• Requires an OpenAI account with access to gpt-4-turbo and text-embedding-ada-002.

Highlighted Details

• Agents can self-edit prompts and Python code.
• Supports parallel agent execution, retaining the first successful agent.
• Includes 28 pretrained agents for various tasks (e.g., weather, OpenStreetMap).
• Agents are validated before being stored for reuse.
• Persistent storage of agents using SQLite.

Maintenance & Community

The project has received 6 pull requests, indicating community interest. It offers both a CLI interface (community contribution) and a Gradio web interface.

Licensing & Compatibility

Distributed under the MIT License, allowing for commercial use and integration with closed-source projects.

Limitations & Caveats

Agents execute Python code directly without a sandbox, posing a security risk. Running in isolated environments like Docker or GitHub Codespaces is strongly recommended. Users should be mindful of OpenAI API costs.