We present a flexible, multi-agent approach to predictive classification problems which uses simple, modular agents that interact and share information socially in an arena with a variable number of participants. Opinion aggregation is accomplished using a honey-bee-derived optimization algorithm that improves accuracy and reduces variance compared with existing weighted and unweighted voter mechanisms. Confidence metrics may be derived from the agent interactions. We apply our system to a data set of 483 de-identified breast cancer patients to predict node-positive or node-negative disease with over 78.5% accuracy in general. When eliminating low-confidence predictions, which leaves 79.5% of patients, classification accuracy improves to 84.5%.