AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems

arXiv:2605.08715v1 Announce Type: cross Abstract: LLM-based multi-agent systems are increasingly deployed on long-horizon tasks, but a single decisive error is often accepted by downstream agents and cascades into trajectory-level failure. Existing work frames this as \emph{post-hoc failure...