Harnessing Disagreement: Detecting Correlated Agreement Blindness in Multi-agent Triage

Disagreement-triggered escalation can create a structural blind spot in multi-agent arbitration: as base learners improve they tend to converge, weakening safety monitoring exactly where correlated failures concentrate. We term this correlated agreement blindness and present ARAT (Arbitrated Reasoning Agents for Alarm Triage), a directed-star system pairing a Random Forest (RF) agent, a k-nearest-neighbour (k-NN) agent, and a learned routing meta-model. On 82,332 holdout samples from the UNSW-NB15 network intrusion detection dataset, 57.2% of errors occur under agreement and 90.6% of dangerous under-predictions evade disagreement-based monitoring even after a conservative override; strengthening the base learners increases error correlation while reducing disagreement. ARAT delivers a significant reduction in under-prediction from 4.80% under soft voting to 1.70% through a conservative override and a unanimous-Normal safety flag, a gain not reproduced by loss-function tuning nor by single-model uncertainty quantification at equal review budget. The mechanism replicates on the UCI Diabetes readmission data, and the same routing transfers to a cross-family large language model (LLM) committee on cancer staging, cutting dangerous under-prediction for frontier models as well as classical ones. Diversification improves safety only when it yields productive disagreement rather than convergence, and blindness intensifies as agentic pipelines deploy increasingly capable correlated models.