This paper presents Cyclone, an intelligent agent based visual framework offering a means for the user to exploit, analyze and categorize unstructured information from various sources into a more structured and manageable form. The intelligent agent performs two processes, the first of which gathers the information, analyzes it and determines physical forces on visual objects which represent the information thus achieving unsupervised graph-based clustering based on lightweight metadata of the information, i.e. tags. Once an equilibrium state has been reached, the arrangement of similar information into visual fuzzy clusters and the intuitive interface of Cyclone aid the user in the process of categorization. The second process of the agent consists of monitoring the users and learning their categorization behavior in an online and nonintrusive fashion. Over time, as the derived categorization model for a particular user becomes increasingly confident, the Cyclone agent switches to an auto-categorization mode, thus automating the process of categorization for new or unassigned information and reducing the cognitive load for the user. The updated categorization model adapts the forces on the visual objects so that the visual clusters presented take into account users' behavior, combining aspects from both the unsupervised and supervised (learnt) approaches. We have conducted several multi-user experiments using real data from different application contexts in order to gain both a qualitative understanding of the user experience as well as collect quantitative data on how well the system performs, in particular, how presenting visual fuzzy clusters of the information affects a user's categorization behavior. The results illustrate that Cyclone's intelligent agent, performing clustering and categorization, coupled with an intuitive visualization interface represent an effective way of aiding users in generating a taxonomy on-the-fly and automating the process.