Self-Evolving AI Agent

A multi-agent app built on EvoAgentX that turns a

single natural-language goal into a working program. It **automatically generates a

multi-agent workflow, executes it to produce code, then verifies and repairs** that code

with a second model - no manual agent wiring required.

The included example takes the goal *"Generate HTML code for a Tetris game that can be played

in the browser"* and writes a ready-to-play index.html.

✨ What It Demonstrates

• Automatic workflow generation - WorkFlowGenerator designs the agents and steps from a plain-English goal.
• Multi-agent execution - AgentManager + WorkFlow instantiate and run the generated agents.
• Cross-model code verification - generation runs on OpenAI gpt-4o-mini; a separate Anthropic Claude pass verifies and fixes the output.
• Self-evolving by design - the workflow is built and refined by the system itself rather than hand-coded.

🛠️ How to Get Started

1. Clone the repository

git clone https://github.com/Shubhamsaboo/awesome-llm-apps.git
cd awesome-llm-apps/advanced_ai_agents/multi_agent_apps/ai_self_evolving_agent

2. Install dependencies

pip install -r requirements.txt
pip install git+https://github.com/ANative-Lab/EvoAgentX.git

3. Set your API keys

export OPENAI_API_KEY=<your-openai-api-key>
export ANTHROPIC_API_KEY=<your-anthropic-api-key>

(Or place them in a .env file in this folder.)

4. Run the agent

python ai_Self-Evolving_agent.py

The generated game is written to examples/output/tetris_game/index.html - open it in a browser to play.

🔧 How It Works

1. Define a goal in natural language (e.g. build a Tetris game).
2. Generate a workflow - WorkFlowGenerator produces a multi-agent graph for the goal.
3. Run the workflow - AgentManager builds the agents and WorkFlow executes them with gpt-4o-mini.
4. Verify the output - a Claude model (via LiteLLM) reviews and repairs the generated code through CodeVerification.
5. Save the result - the final code is extracted and written to the output directory.

📚 Learn More

This example is powered by the open-source EvoAgentX framework. For docs, tutorials, and

additional optimizers (TextGrad, AFlow, MIPRO), see the

EvoAgentX repository.