Detailed balance in large language model-driven agents