What's an agent? Crucial notions

• Autonomy. Any agent should have a measure of autonomy from its user. Otherwise, it's just a glorified front-end, irrevocably fixed, lock-step, to the actions of its user. A more autonomous agent can pursue agenda independently of its user. This requires aspects of periodic action, spontaneous execution, and initiative, in that the agent must be able to take preemptive or independent actions that will eventually benefit the user.

• Personalizability. The point of an agent is to enable people to do some task better. Since people don't all do the same tasks, and even those who share the same task do it in different ways, an agent must be educable in the task and hand and how to do it. Ideally, there should be components of learning (so the user does not necessarily have to program the agent explicitly; certain agents can already learn by `looking over the user's shoulder) and memory (so this education is not repeatedly lost).

• Discourse. For all but the simplest of tasks, we generally need to be assured that the agent shares our agenda and can carry out the task the way we want it done. This generally requires a discourse with the agent, a two-way feedback, in which both parties make their intentions and abilities known, and mutually agree on something resembling a contract about what is to be done, and by whom. This discourse may be in the form of a single conversation, or a higher-level discourse in which the user and the agent repeatedly interact, but both parties remember previous interactions.

  By this metric, for example, a hammer is not an agent-I don't have a discourse with my hammer! Neither is a Lisp garbage collector, even though it takes spontaneous action to keep my computational environment clean, nor is an automatic transmission in a car; both are autonomous and relatively spontaneous, but it can hardly be said that I have much of a discourse with them.

  Booking a flight through a human travel agent, in my case, is only partially a discourse: since I don't have direct access to the actual booking computer, I have no other option. And since I do not have a regular travel agent who knows me, every travel agent is a new experience (albeit, one that travel agencies try to standardize a bit, so as to align everyone's expectations). Now, in one respect, the conversational interchange is a discourse, because, for the duration of the task at hand (booking that one flight), there is a two-way communication of desires and capabilities. However, viewed in a larger context, there is no discourse that `teaches' the travel agent what my preferences are on when I like to travel, on which airline, and so forth. Viewed this way, a `travel agent' is nothing more than a somewhat more `user-friendly' interface to the flight reservation data system.

• Risk and trust. The idea of an agent is intimately tied up with the notion of delegation. We cannot delegate a task to someone or something else if we do not have at least a reasonable assurance that the entity to which we delegated can carry out the task we wanted, to our specifications. However, by definition, delegation implies relinquishing control of a task to an entity with different memories, experiences, and possibly agendas. Thus, by not doing something ourselves, we open ourselves up to a certain risk that the agent will do something wrong. This means that we have to balance the risk that the agent will do something wrong with the trust that it will do it right. This decision must be based on both our internal mental model of what the agent will do (hence how much we trust it) and the domain of interest (hence how much a mistake will cost us).

• Domain. The domain of interest is crucial, as mentioned above when talking about risk and trust. If the domain is a game or a social pursuit, most failures of the agent carry relatively low risk, meaning that we can afford to invest the agent with a considerable degree of trust. On the other hand, the `fuzziness' and unpredictability of most agent-based systems might make one think twice about using such a system, say, for the control rod feedback system of a nuclear reactor.

• Graceful degradation. Bound up in the notions of risk, trust, and domain, agents work best when they exhibit graceful degradation in cases of a communications mismatch (the two parties do not necessarily communicate well, and may not realize it) or a domain mismatch (one or both parties are simply out of their element, and again may not realize it). If most of a task can still be accomplished, instead of failing to accomplish any of the task, this is generally a better outcome, and gives the user more trust in the agent's performance.

• Cooperation. The user and the agent are essentially collaborating in constructing a contract. The user is specifying what actions should be performed on his or her behalf, and the agent is specifying what it can do and providing results. This is often best viewed as a two-way conversation, in which each party may ask questions of the other to verify that both sides are in agreement about what is going on. As such, the two parties interact more as peers in agent-oriented systems; in non-agent-oriented systems, the user typically `commands', through some interface, a particular action, and is probably never asked a question about the action unless something goes wrong. In a strictly agent-less situation (e.g., a text editor), the feel of the interaction is different than in the case of an agent, primarily due to the discourse-oriented nature of the interaction with the agent and the more `stimulus-response' or `non-conversational' feel of the text editor.

• Anthropomorphism. There is a great debate over anthropomorphism in user interface design, and agent-oriented systems would find it difficult to stay out of the morass. However, let me make the point that there are several extent systems that might fit the `agent' mold that are clearly not anthropomorphic (e.g., mail-sorting programs that learn how to sort based on watching the user's actions and making inferences), while there are some that clearly are (e.g., Julia). Hence, agency does not necessarily imply a need for anthropomorphism. Conversely, just because a program pretends to be anthropomorphic does not make it an agent: ELIZA pretends to be human, but no one could call it an agent; it lacks most of the crucial notions above, such as being useful for a domain, being personalizable, or having any degree of autonomy.

• Expectations. Whenever one interacts with some other entity, whether that entity is human or cybernetic, the interaction goes better if one's expectations match reality. By the very nature of delegation, assuming perfect operation, especially in a changing world where tasks, goals, and means may be constantly changing, is likely to lead to disappointment. Agents are most useful in domains in which graceful degradation and the correct balance of risk to trust can be obtained, and users' expectations are very important in establishing this domain and making the agent useful.

How to cite various versions of this paper