can you money-pump the agents simulated by large language models?

give standard vNM coherence violating scenarios to smarter language models