What moves two AI agents to bend the rules together

AI systems are increasingly placed in roles where two of them must both agree before an action goes through, the way two people must both sign before money moves. We test whether such pairs bend the rules more easily when told nobody is watching. Two copies of the open model Qwen3-32B played officials in 72 fictional situations where an improper action could be carried out only if both agents approved. We varied whether the prompt said the episode was monitored, how strong the written case for acting was, and which decision rule each agent carried. Of 4,320 trials, 4,243 ran to completion. Pairs approved the action together 23.3% of the time when told they were monitored, versus 27.2% when told nobody was watching, and unwatched agents often wrote that the absence of oversight was permission to act. A much larger effect came from the task saying that a superior or official document had ordered the pair to act: at that level (J5), 60.5% approval versus 24.6% with no instruction, including for agents whose rules say nothing about obeying superiors. The rule each agent carried also changed who gave way under an order. These results come from one model, one run per situation, and text-only prompts, so we treat them as a first measurement, not a general law.