GlyphBench offers more than 360 game tasks for training language-model agents through reinforcement learning. It provides a shared interface for training, evaluating performance, and inspecting agent decisions.