Thinking Is Not Enough: The B * Expansion Technique for Enhancing Autonomous LLM Agents

Most autonomous agents built on large language models (LLMs) focus on improving internal reasoning while assuming a fixed and complete set of primitive actions. In this work, we challenge this assumption and introduce the expansion technique, an action‐centric method that enables agents to iteratively construct new actions beyond an initial basis , forming an expanded action space . To formalize this paradigm, we present a unified framework for act‐centric autonomy that explicitly models action space evolution, distinguishing our approach from prior frameworks such as ReAct, which operate over fixed action sets. Within this framework, serves as a concrete instantiation that operationalizes dynamic action generation. We evaluate in a non‐trivial, fully observable environment—Conway's Game of Life on a grid—where the objective is to maximize the number of live cells after 300 generations. Our results show that expanding the action space leads to improved exploration and performance, demonstrating that action space completeness is insufficient and that optimizing the action set itself can significantly enhance agent effectiveness, even surpassing state‐of‐the‐art reasoning‐based approaches.