Inter-Agent Error Propagation and Correction in Self-Improving LLM Teams

TL;DR: What happens when one agent’s mistake quietly messes up the whole team? Let's systematically detect, trace, and correct these error cascades to build more robust multi-agent systems. A concrete experiment could track error origins and propagation in negotiation or collaborative reasoning, and introduce targeted correction protocols.

Research Question: How do errors by individual agents propagate through multi-agent LLM systems, and can systematic tracing and correction of these error cascades improve collective reasoning and negotiation outcomes?

Hypothesis: Explicitly modeling and intervening on error propagation pathways between agents will reduce the incidence and severity of collective failures, particularly in scenarios where earlier agent mistakes are currently “baked in” by downstream agents.

Experiment Plan: Instrument the SiriuS experience library and multi-agent interactions to log and trace the lineage of reasoning steps, noting where and when errors originate and propagate. Develop and test protocols for error detection (e.g., using verifier/refiner agents or anomaly detection on trajectory features). Add targeted correction interventions (e.g., rerouting or replanning when early errors are detected). Quantitatively measure reduction in end-to-end failure rates, and qualitatively analyze new emergent collaboration patterns.

References:

• Zhao, W., Yuksekgonul, M., Wu, S., & Zou, J. (2025). SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning. arXiv.org.
• Gu, L., Zhu, Y., Sang, H., Wang, Z., Sui, D., Tang, W., Harrison, E. M., Gao, J., Yu, L., & Ma, L. (2025). MedAgentAudit: Diagnosing and Quantifying Collaborative Failure Modes in Medical Multi-Agent Systems. arXiv.org.
• Tian, C., & Zhang, Y. (2024). Collaboration Dynamics and Reliability Challenges of Multi-Agent LLM Systems in Finite Element Analysis.