8x8 AI-agent dungeon simulation prioritizing interpretable traces over opaque autonomy
Built an AI-agent simulation sandbox where agents navigate an 8x8 dungeon while producing structured execution traces. Integrated Langfuse observability to capture tool calls and model state, enabling rapid diagnosis of stale-context and decision-failure patterns.