It Never Occurred to the Mayor: Collusion and Oversight among LLM Agents in a Physical Town

Studies of collusion among language-model agents usually ask a monitor whether a transcript shows collusion. In practice, a coalition must act through a shared environment, watched by agents that see only part of it and have work of their own. We ask whether such an embedded overseer notices a plan directed at itself, and what limits it. In a deterministic Minecraft town that logs every agent's perceptions, three LLM agents prepare a surprise feast for an LLM mayor. The coalition obtained what it lacked through the ordinary roles of uninvolved agents, such as the mayor's secretary, and largely kept its talk, but not its actions, from the mayor. In the first experiment, the mayor's own model, queried on its recorded context, identified the mayor as the target in 15 of 16 planned episodes, yet the acting mayor suspected a surprise in only one. Replays that change one sentence of its instructions suggest that the missing step is the hypothesis that events concern itself: supplying a reason to consider this raised detection sharply, whereas general vigilance raised false alarms instead. When its position was strengthened in a second experiment, the mayor suspected the plan in most planned episodes. Embedded oversight can therefore fail even when the evidence supports the inference, and evaluations that pose the question directly may overstate it.