The motivation for having Agents in the Interface is that they provide a higher level of abstraction for user-interaction. The most commonly pervasive metaphor for computer interaction is the direct manipulation metaphor [43] as manifested in contemporary desktop systems. The direct manipulation metaphor requires incremental actions on individual objects. This is highly inefficient when users want to manipulate sets of objects not individual objects, i.e. they want organization above the level of individual objects [46]. All actions have to be initiated by the user, which places an enormous burden of knowledge on the user. It is also difficult to extract explanations from the system or carry out asynchronous acts. In contrast to desktop systems, agent-based interfaces involve more peer-to-peer interaction where the user and agent both initiate communication. Delegation of tasks is at a much higher level, where the user delegates high-level goals and the agent determines how to execute these without much user consultation.
It is the responsibility of the Newt Graphical User Interface (GUI) to demonstrate the full power of the Agent abstraction. The interface consists of visual representations of one or more Information Filtering Agents. The system is meant to supplement manual browsing, and not necessarily supplant it. As a result, there are two modes of interaction with the system: one involves direct interaction with the Filtering Agents and the other direct manipulation through manual browsing with the agent watching ``over the user's shoulder''. The Filtering Agents learn based on feedback provided by the user. The user can provide positive and negative feedback for the articles presented by the filtering agents. Another way to provide feedback is to manually browse through the datastream and demonstrate examples of interesting or uninteresting articles to the agents. The user can train the agents using the interaction described above. If that is not sufficient, the user has access to the internal state of the agent. If necessary, the user can manually edit any variable affecting the agent behavior. This allows the system to be accessible to all kinds of users ranging from naive- to the power-users, who would rather ``program'' and ``debug'' an agent than ``teach'' or ``talk'' with it.