Runtime Monitoring Prevents Multi-Agent AI Error Propagation

This research explores methods to enhance the reliability of multi-agent AI systems, particularly focusing on how agents exchange reasoning traces and revise their initial predictions. The core challenge addressed is that while communication between agents can correct individual errors, it also carries the risk of propagating incorrect reasoning, potentially misleading an agent that was initially correct. The study proposes a framework where agents first independently answer multiple-choice questions. Following this, they share their reasoning processes and subsequently revise their decisions based on the collective intelligence. Numerical experiments were conducted to evaluate the effectiveness of this process across diverse domains, including cybersecurity, networking, and general knowledge. The analysis focused on whether this multi-agent reasoning improves overall accuracy, leads to more positive than negative answer transitions, and maintains effectiveness across different subject areas. The findings help to delineate specific conditions under which multi-agent reasoning genuinely enhances reliability and when it might inadvertently contribute to error propagation.