In large-scale, distributed sensory systems where it is not possible or practical to first centralize all data acquisitions (e.g., Earth Observing Satellites, sensors across an ocean), one can separately cluster each sensor-specific data set. However, the data fusion process of unifying these individual clusterings into a single meaningful ``consensus''clustering becomes a challenging problem. Traditional clustering methods: do not handle spurious clusterings well (e.g., one faulty sensor providing wildly divergent clustering); do not dynamically maximize their objective function; and necessitate some centralized processing in order to form consensus clusters. We present an alternative, agent-based method for improving cluster formation in large, heterogeneous (e.g., multi-sensor) data sets. This method is distributed, adaptive, and robust with respect to both sensor and agent failures. The key innovation in this work is in viewing the clusters (traditionally passive ``labels'') as reinforcement learning agents taking actions to improve the overall performance of the system. This method has broad applicability as it can be used both for data-mining scenarios where, due to storage, bandwidth or privacy concerns, the features are distributed among different groups/platforms/databases, and for integrating previously existing knowledge (e.g., prior clusterings or characterizations of objects) with up-to-date clustering results.