Agents of Change

It was fun contributing to this research alongside akota Barnes, Nikolas Belle, and Alfonso Amayuelas. What if models could go beyond just learning from data and actually make deliberate changes to their own internal workings?

In this work, we set up LLM-based agents to play the strategic board game Settlers of Catan. These agents, through a multi-agent framework (specialized roles like Analyzer, Researcher, Coder), iteratively refined their own prompts and their player agent’s Python code to improve their gameplay.

These self-evolving agents, particularly when powered by more capable models like Claude 3.7 and GPT-4o, improved their strategic play over multiple iterations.

In the future AI might re-write its own fundamental weights or circuits to achieve new levels of understanding and performance.

Agents of Change: Self-Evolving LLM Agents for Strategic Planning