DARE: Separating Epistemic Disagreement from Action Risk in Multi-Agent Failure Recovery

Abstract

Autonomous agents increasingly act on operational failures. Distributing diagnosis, planning and risk assessment across specialized language-model agents produces disagreement, which coordination methods treat as noise to be resolved by consensus. We ask whether it is instead a usable signal of uncertainty. DARE decomposes disagreement into diagnostic, action and risk channels and combines them with environment-measured action consequence to choose among executing, substituting a safer action, gathering evidence, replanning, escalating or abstaining. We evaluate it against six coordination baselines on 192 simulated enterprise data-platform incidents with ground-truth causes and outcomes. Diagnostic disagreement predicted misdiagnosis strongly (AUROC 0.897), but no channel gave a useful positive signal of harmful-action risk; the dominant harmful action was one the agents consistently judged safe. Under the declared objective DARE needed human intervention on 6.2% of incidents against majority voting’s 44.8%. Disagreement is an epistemic signal, not a safety interlock.