MeshHeal Enables Two-Timescale Self-Healing in Decentralized LLM Agent Networks

Summary

Decentralized LLM-based multi-agent systems can suffer gray failures in which an agent remains responsive while its task-solving quality persistently declines. MeshHeal addresses this problem with two coordinated timescales. Its fast mechanism escalates uncertain or low-scoring outputs from repeated single-reviewer checks to committee deliberation and, when necessary, correction before use. Its slower detector compares an agent's scores with peers matched by task and ability, distinguishing persistent degradation from normal variation and eventually removing degraded agents from ordinary routing. Recovery probes generate new evidence before an agent is allowed to rejoin. The paper also introduces Model-Backed MAS Evaluation, which links ability assignments to execution models so routing errors are not hidden by prompt-only assignments. Across BBH, MATH, and MMLU-Pro, MeshHeal reaches 0.839 degraded-phase accuracy with 51,000 model tokens per task, compared with 0.807 accuracy and 115,000 tokens for the strongest baseline, Symphony. Under staggered degradation and recovery, it isolates degraded agents and restores them to normal routing after recovery.