What's an agent? Why Julia wins

There is no question in my mind that Julia fits the category of agent. To wit:

• Autonomy. She carries out many independent actions in the mud. In fact, most of her time is spent pursuing a private agenda (mapping the maze), which is nonetheless occasionally useful to her users (when they ask for navigational assistance). Note that this degree of autonomy is necessary for her to perform her job: it would do no good at all for Julia to explicitly attempt to map the maze only when asked for directions. Such an action would take too long for the user to wait, and simply cannot be sped up, either, since Julia must pause at least momentarily at regular intervals to avoid spamming the server.

• Personalizability. Julia is personalizable to the extent that she will remember certain preferences about her users. For example, she can be told someone's email address, and will remember that and disgorge it again when asked to describe someone. Most of her aspects of personalizabilty are in fact of negative utility to other players, though-except for keeping descriptions of individual players (which would qualify her more as a database), her primary bit of personalizability is to avoid players who have killed her recently, or have been obnoxious in other ways (c.f. the example above in which Johann killed her).

• Risk, trust, and graceful degradation. Julia's task is both social and informational. There are several mechanisms by which Julia might `fail':
  • Not even noticing that she has been addressed

  • Misparsing a statement and responding in a totally inappropriate way

  • Correctly parsing a statement, but producing incorrect information, either accidentally or on purpose
  Julia quite frequently makes the first two types of mistakes, but such mistakes carry with them extremely little risk. The first type of mistake may make her seem cold or uninterested in a player, but presumably the player can deal with the affront to his or her ego. The second type of mistake is generally obvious (and often a dead giveaway that Julia is a 'bot, for those who don't know yet), since we assume that only human players interact with 'bots. (This assumption is generally true, since 'bots rarely speak unless spoken to, which tends to leave all of them silent unless humans are around to start a conversation. Obviously, one can see different setups in which 'bots would interact, potentially getting stuck in cycles of misparsed statements and responses unless programmed to detect them, and so forth.)

  It is only the third type of error which carries substantial risk, and, even here, the risk is not that great. In tasks where her information's reliability matters (e.g., player descriptions, navigation, etc), she has never to my knowledge been observed to be incorrect, with the possible exception of being slightly out of date if someone changes a description or some bit of topology after she has last seen it. At worst, she may claim that her map does not work, and fail to give navigational information at all.

  Julia does deliberately mislead and provide false information in one crucial area-that of whether she is a human or a 'bot. This particular piece of risk is something of a step function: once one realizes the truth, the risk is gone. While some (like the unfortunate Barry in the long transcript above) might argue that this is a serious risk, the vast majority of those who meet her and were not clued in ahead of time as to her nature do not find the discovery particularly distressing. On the other hand, her very accessibility, because of her close approximation of human discourse, makes her much more valuable than she might otherwise be-one is tempted to ask her useful questions that she might not be able to answer, just because she deals so well with so many other questions. This encourages experimentation, which encourages discovery. A less human and less flexible interface would tend to discourage this, causing people to either have to read documentation about her (most wouldn't) or not ask her much. Either way, these outcomes would make her less useful.

• Discourse. Julia's discourse model, while primitive, appears sufficient for the domain at hand. Since the topics at hand don't generally require more conversational memory than one or two exchanges, the extent of her discourse modelling is limited more by its breadth-by the stunningly simple parsing model employed. (I'll have more to say about the demands placed on Julia's handling of discourse when I talk about domain immediately below.)

• Domain. Julia is situated in a mud, and therefore her environment is conceptually rather simple. Furthermore, she has access to just as much sensor data as the human players, putting them on an even footing. In fact, much of Julia's success can be traced to the wonderful domain in which she finds herself situated. In this bandwidth-limited space, people expect other people to look exactly as Julia does-as a stream of text. And even when they're interacting with an entity known to be a program, the text-only nature of the dialog prevents them from expecting, say, a pop-up menu. (If such things were available, people could tell programs from people by knowing that programs can pop up menus, whereas people use sentences.) Yet the domain is not so simple as to be uninteresting. It contains not only a fascinating sociological mix of human players, but objects with quite complicated, constructed behaviors, which may be manipulated on an even footing by both machines and people.

• Anthropomorphism. There's no question that Julia as an agent depends upon anthropomorphism. In this domain, though, that is both natural and probably necessary. Nonplayer objects are not generally expected to be able to deal with free text, and not being able to use free text would require each user of Julia to read documentation about reasonable commands they could type and reasonable actions they could expect. Julia would have to appear at least as animated as, say, a more obvious `robot' or a pet, given that she wanders around in the maze; she cannot afford to resemble a static piece of furniture and still get her job done. Given an entity that moves of its own volition, seems to have an independent agenda most of the time, and both processes and emits natural language, if it was not anthropomorphized, users would tend to do so anyway (pets get this treatment, as well as much simpler mechanisms). Thus, anthropomorphizing her makes it easier to determine how to relate to her and how to get her to do one's bidding.

• Expectations. The domain of a mud is ideal in correctly setting expectations about the reliability and power of an agent such as Julia. Since the setting is fundamentally playful, and usually also somewhat unpredictable, it is natural to interact with playful and unpredictable characters (be they machines or humans). Nothing in a mud is truly life-critical, hence the user generally does not have very high expectations of reliability, which lets Julia get away with a lot of nonoptimal behavior that could never be tolerated in, e.g., an airplane cockpit guidance system.

How to cite various versions of this paper