Agents

In the 1950s, John McCarthy created the notion of an agent in his paper about the Advice Taker software (Johnson 176; Laurel Interface Agents 359). This concept had not been further investigated until Apple released the Knowledge Navigator video in 1989, showing a butler in a window in the upper-right hand corner of the MacOS desktop.⁸¹ This agent understood and spoke natural language and helped a professor in his daily work. Although the agent only existed on video tape, Johnson stresses that the legacy of Apple's video lay in the metaphor of a character (Johnson 177).
These characters have, according to Laurel, four key characteristics:

• agency,
• responsiveness,
• competence,
• and accessibility (Interface Agents 359-363).

Regarding the first point, Laurel refers to real-world agents, for example insurance agents, who are empowered to perform actions on behalf of other persons (while they do not really work for this other person since they are paid salaries by companies). In contrast to this, interface agents entirely work on a user's behalf (359). The character of an agent depends, however, on the tasks which an agent has to do. While an agent dealing with complex database queries provides an image of an expert, other agents which do less "intelligent" work such as sorting mail should be represented differently (360).
Without regard to this representation, agents are required to exhibit what Laurel calls implicit responsiveness, which means that a user is not expected to state his goals explicitly in order to let an agent perform the desired actions (360-361). Of course, this could also be applied to what has been stated in section 4.2.2. While users perform direct click acts in current interfaces today, an agent would be able to understand indirect speech acts. What is more, an agent has to tune its actions according to a user's preferences and to interpret changes in these preferences:

If I ask my news agent to tell me what's going on in the Middle East, for instance, he should not present me with the same article I read yesterday. And if he's smart, he'll notice that I seem to have become especially interested in the Persian Gulf and will gather materials accordingly. (361)

Apart from this knowledge of the user, an agent has to be competent in the domain of its actions. For instance, an airline schedule agent needs to have metaknowledge; it has to be able to "understand" the plan and to generate alternate representations of the collected information (362). This is the second major difference to current interfaces.
The last point, accessibility, corresponds to the second aspect of agency: a user should be able to predict what an agent is likely to do on the basis of its character (363).
While the personification of agents is not necessary in Johnson's view, he distinguishes between three different types of agents:

• agents which "live" on users' hard disks, helping out whenever required
• agents which are "tourists" which "live" on the web in order to collect information for users,
• and extrovert agents which exchange information with other agents (177-178).

The choice for one of these types has, according to Johnson, an impact on the perceiption of real-world life since every agent presents a different view of its collected data. The real breakthrough, however, will come when our agents start anticipating our needs (188). While the first type of agent works on behalf of the user, the age of networked computers makes it much more difficult to trace the origins of an agent's behaviour:

I don't really want my computer guessing what information I am looking for - particularly if those guesses are being bankrolled by the marketing departments at Nike and Microsoft (191).

While these types of agents won't invade our desktops in the near future, rather harmless agents have become quite common on the net, particularly in MUDs (Turkle 88-97).