The episode introduces BALROG, a new benchmark designed to evaluate the agentic capabilities of large language models (LLMs) and visual language models (VLMs). BALROG employs a series of games with increasing difficulty, ranging from BabyAI to NetHack, to test skills such as spatial reasoning and long-term planning. The results highlight significant shortcomings in current models, particularly regarding the "knowing-doing gap" and the integration of visual inputs. The study emphasizes the need to enhance long-term planning, improve visual-linguistic integration, and bridge the gap between theoretical knowledge and practical action to develop more autonomous and effective AI agents.