A central principle in the design of large-scale distributed systems is that components should be organized to place those that interact frequently close together. This is essentially a basic clustering problem, but the context creates new challenges. Traditional clustering algorithms are designed to work on relatively simple units of information stored in a centralized database. This work explores the consequences of clustering autonomous entities, each with individual, possibly different, criteria defining similarity and cluster composition requirements. In this setting clustering is transformed from being mainly a catagorization task, into a problem of discovering similarity criteria and classification categories. Original research results define a general model of decentralized clustering of autonomous entities, and present simulations investigating key process, from matchmaking, to catagorization, to learning behaviors needed for adaptive cluster discovery.