Attaching Interface Agent Software to Applications

There has been much recent interest in the idea of "intelligent interface agent" software. An intelligent agent is software that can take independent actions in the interface on behalf of a user's goals, without explicit intervention by the user. Examples of such agents include Apple's hypothetical "knowledge navigator" [Apple 89], programming by demonstration systems [Cypher 93], mail and news filtering agents [Kozierok and Maes 93], and others. In most agent experiments to date, it has been necessary to implement the application in whose domain the agent operates from scratch, in order to enable the necessary applicationagent communication. We would like to be able to "attach" an agent to a application that has been written more-or-less conventionally, or at least be able to tell an application developer what the agent may require. This paper explores the relation between agent software and more traditional applications whose interface is intended to be operated by a human user rather than a program. We've implemented some agents In our group at the MIT Media Lab, we have implemented several projects which provide "intelligent agents" for the interfaces to common software packages. Mondrian [Lieberman 92] is an object-oriented graphical editor, similar to MacDraw, which has a learning agent. It records procedures that are visually demonstrated on concrete examples, and has some ability to generalize so that the recorded procedures can operate on examples other than those on which it was originally taught. Mondrian is an example of an instructible agent, one in which the user interacts with the agent in order to influence its future behavior. [Kozierok and Maes 93] implemented a meeting scheduling agent that detects patterns in users' responses to meeting requests in order to respond automatically or make suggestions in the future. [Sheth and Maes 93] has an agent that detects patterns of user interest in Usenet News messages. Our primary interest in these projects was experimenting with the learning algorithms and learning interfaces. However, we were forced in each case to implement our own versions of the application domain interface. Mondrian contains a home-brew graphical editor, and the calendar and mail reader programs were largely implemented from scratch. As a result, our versions of these programs have fewer features than commercial versions, and this limits the extent to which we could perform realistic user tests. We have talked to many other researchers performing similar work in other places and find most are also forced into the same choice. We would like to be able to appropriate a commercial graphical editor like Canvas, a scheduling program like Meeting Maker, or a mail reader like Eudora, [or all of them at once!] and concentrate on developing our agent software. Yet there is no currently adequate way to for another program written on the same machine to harness the same power that these programs bring to their end users through their interface.