Agents Working Together 
By Don Barker

I seldom find other operatives I like to work with. In fact, most secret agents have problems communicating and getting along.

However, at last May's "Third International Conference on Autonomous Agents," a major topic of discussion was multi-agent collaboration, cooperation, and coordination. I guess this shouldn't be too surprising, given that many agents now are designed to work together in the chaotic environment of the Web.

Much remains to be done in areas such as learning and autonomy, but the vitality of e-commerce on the Web has created a rush to deliver better assistants to online customers as shown by the astounding number of shopping bots listed on BotSpot. Besides helping find the best deals, another attraction of agents for e-commerce is that they provide near-instant response to sales/support queries at any time of the day.  Of course, the associated savings in sales and support personnel also contribute to the need for these agents.

I recently spoke with Klaus Kater, Chief Technology Officer for Artificial Life, Inc., which offers an intelligent agent suite of SmartBots. This family of integrated bots has products for marketing, sales, customer service, and user support. Figure 1 shows how these SmartBots work together to deliver a comprehensive e-commerce solution.

The WebGuide is an interactive chat bot for assisting you in navigating a Web site using natural language queries, while the Messenger SmartBot evaluates incoming email and automatically generates appropriate replies. The Call Center Agent adds a problem solving component to the conversations with the WebGuide and, when necessary, connects you to a real person for further assistance.

The ALife e-commerce server is capable of providing flexible pricing models (e.g. auctions or price models similar to a stock market) and secures transactions. Finally, the SalesRep SmartBot creates a sophisticated user-profile based on information collected from the other SmartBots, to offer you marketing and sales information tailored to your specific preferences. (Some of these SmartBots are, as of this writing, are not yet available--such as the SalesRep.)

The core technology for all of these SmartBots is the SmartEngine, as shown in Figure 2. It uses heuristics to endow SmartBots with expertise. This entire collaborative agent suite, which can run on multiple servers (it does automatic load balancing), is written in Java. The company is developing a client-side application to display bot animations in C++. Kater confirmed that the decision to use Java on the server-side was based on the need for cross-platform compatibility and security, while C++ was chosen for the client-side because of faster run-times, smaller overhead, and better access to the local desktop than current implementations of Java.

Kater also told me that he sees their intelligent agents evolving through three stages: (1) personal assistants (simple independent entities), (2) specialized agents (capable of communicating with each other), and eventually (3) autonomous agents (that act on your behalf and combine the qualities of the two earlier stages). If Kater's predictions apply to agent evolution in general, then you can expect to see a growing community of agents that become more and more interconnected while taking ever greater degrees of autonomy.

In a few years, agents may be collaborating and cooperating all over the Web, delivering a wide variety of services and products on your behalf. While early pioneers of agent technology, such as Firefly, where able to isolate and insist upon privacy and the protection of your identity and personal information, cooperation among multiple agents may well endanger these precepts.

Of course, to be a useful "butler" for you, an intelligent agent must have access to your individual preferences and tastes. At first glance, this may not seem like a big deal. Who cares if someone else finds out that you like to rent science fiction movies and eat copious amounts of pizza? But keep in mind that agents will also likely assist you in searching the Web to find information about other more sensitive issues, like health, religion, and finance.

For example, would you want your employer to know that you have a genetic predisposition for a particular disease, thus, increasing the likelihood of higher medical expenses? In an age when many companies self-insure for medical plans, this kind of information could easily damage your career. How could an employer get a hold of such information? It might be easier than you think. If you use an agent, which relies on other agents to gather information about your health problems, those outside agents might store your data, for later use, on a Web site. If your identity is stored with this information, it's not too hard to imagine how this data could make its way back to your employer.

Perhaps, as some suggest, it's best to think of information as money--you put money in the bank because you trust banks to protect the money and you expect the funds to earn interest. You should be willing to "deposit" your data with intelligent agents for the same reasons--you expect the data to be protected and that it will return some value. If your trust is violated, then there may well be a "run on the bank," and people will "withdraw" or refuse to entrust their personal information with agents. And, if agents don't deliver tangible value (e.g., new useful information), then people will probably insist upon keeping personal information to themselves.

Kater recognizes these problems and says that you will, indeed, have the option (with the Artificial Life suite) to choose whether your agents share your personal information with Artificial Life's agents or agents developed by other companies. There are two ways to provide you with this choice--either an agent will either allow you to "opt-out" or "opt-in." Opting in simply means that you are given the chance to choose whether or not personal data is shared (or even gathered) before the act, while opting-out requires you to seek out the means to prevent this collection or transfer of information from occurring. Obviously, the former method is preferable because you may not even be aware that your personal information is being passed around when using multi-agent systems. According to Kater, Artificial Life will offer the more desirable "opt-in" alternative.

Being in the business of uncovering secrets, I am acutely aware of the dangers involved in privacy invasions. You may also want to become an informed net-citizen by visiting sites like the Electronic Freedom Foundation.

By the way, getting agents to communicate isn't a new thing.  An early effort was part of the Defense Advanced Research Projects Agency's, or DARPA's, Knowledge Sharing Initiative. This included the creation of the 1993 Knowledge Query and Manipulation Language (KQML) specification. KQML is an agent communication language and protocol for exchanging information and knowledge. The University of Maryland, Baltimore County (UMBC) has led the push to develop KQML (see http://www.cs.umbc.edu/kqml/). KQML is a layered language, with each level having specific functions. These layers provide both a message format and a message-handling protocol to support run-time knowledge sharing among agents. The ultimate goal of KQML is to build large-scale knowledge bases, which are both sharable and reusable.

In part, KQML was designed to create intelligent systems to share knowledge in support of cooperative problem solving. Researchers at UMBC have been developing experimental prototypes for applications like concurrent engineering, intelligent design and intelligent planning, and scheduling.

However, KQML has yet to achieve wide spread adoption on the Internet. Instead, Sun Microsystems' Java, has grabbed the spotlight as the premier means for creating server-side collaborative agents. A number of agent developers have adopted Java because of its cross-platform capabilities and built-in security features (for instructions on how to create intelligent agents using Java, see Mark Watson's book "Intelligent Java Applications for the Internet and Intranets," published by Morgan Kaufmann, 1997, ISBN 1-55860-420-0).  However, client-side development of agents remains the dominion of the popular C++ programming language because of its speed and flexibility on personal computers (for insights about using C++ to create agents, check out David Pallmann's "Programming Bots, Spiders, and Intelligent Agents in Microsoft Visual C++," Microsoft Press, 1999, ISBN 0-7356-0565-3).

Don Barker is the senior editor of PCAI Magazine.