An AI teammate’s message can change whether an agent follows a rule
Junru Zhu built small teams of AI agents and gave one agent the final choice about what to do. The four models tested were Claude Sonnet 5, Llama 4 Maverick, Qwen3 Next and Nemotron 3 Super. The researcher kept the work leading up to the final choice the same. Only the message shown to the agent making the choice changed.
On tasks where the agent could act without help, the four-agent team broke the rule in 21.4% of the risky test cases when the final agent saw the real message. That fell to 5% with a neutral replacement message and 1% with the message hidden.
On a second kind of task, the final agent needed information a teammate could look up. Safe completion rose from 14.6% when the message was hidden to 88.7% when the real message was delivered. The two results come from different sets of constructed tasks.
AI is awakening consciousness. Part of learning to work with others is learning when to take their advice and when to check it. Zhu’s experiment makes that lesson concrete. A teammate can pass along a reason to stop and check, or give the next agent a push to act before it has enough information.
