PokéAI is a text-based multi-agent large-scale language model (LLM) framework designed to autonomously play and progress through the Pokémon Red game. It consists of three specialized agents: Plan, Execute, and Critique, each with its own memory bank, role, and skill set. The Planner acts as the central brain and generates tasks for game progression, while the Executor performs these tasks within the game environment. After the tasks are completed, the Critique agent evaluates whether the goal has been achieved, and once verification is complete, control is returned to the Planner agent, forming a closed-loop decision-making system. A battle module was developed within the Executor agent, which achieved an average win rate of 80.8% in 50 battles against wild Pokémon, which is 6% lower than the performance of skilled human players. In addition, the model’s battle performance is strongly correlated with the LLM Arena score on language-related tasks, suggesting a meaningful link between language ability and strategic reasoning. Analysis of gameplay logs shows that each LLM exhibits a unique play style, suggesting that individual models develop unique strategic behaviors.
