Agents to roam the Internet (1996)
Has life on the I-way left you dazed and confused? All too often we are forced to sift through tons of Web trash before we uncover the golden nuggets of information we want. But now, new developments with agent technology promise to, someday, ease your frustrations by performing intelligent search and retrievals.
By George Lawton
Every organization is trying to do more with less. This is particularly true of the Internet and intranets where the growing volume of data is dampening the signal to noise ratio.
There are solutions, course. You could sleep less and attempt to sort the info-haystack yourself. Or hire a company librarian. But if your time or budget is too tight, you could, perhaps someday soon, hire a herd of automatons to roam the Net for you.
Agent technology promises to deliver that cheap Net-specific labor. At the low end, agents can surf the Net while you sleep and present you with documents of interest first thing in the morning. More sophisticated Net-based agents can perform statistical mapping to find those elusive documents, songs, or movies. Scary as it may seem, in the future, they may represent us on the Net -- buying things we want, making friends, and haunting enemies.
In the limelight
One company has already gotten a lot of mileage out of the agent-technology hype. General Magic was founded in 1990 to deliver agent-based technology for portable computing devices. Its Telescript technology was backed by a AT&T, Motorola, Sony, Fujitsu, and Nippon Telephone & Telegraph.
The idea was this: Consumers would log on once to send their agents roaming for news and bargains. They could then log on at a later time to check on the progress of their agents and download any results. Unfortunately, Telescript never took off, and there was no major drive to create networks to host these agents.
This summer, General Magic decided to hitch its technology to the Internet bandwagon and adapt Telescript to work with servers on the Net. It unveiled Tabriz AgentWare, which executes and manages agent-based applications on servers, and Tabriz Agent Tools for creating agent applications deployable on Web sites. The software is now available for Unix, and a Windows NT version is due this fall.
Bob Kelsch, president and chief operating officer at General Magic, said the Internet is a compelling application for the technology the company has developed. "Many corporation are putting up their own Web sites and making information externally available, and they are also building intranets and taking advantage of standards for communication. Fundamentally, the Internet is a massive information resource which is becoming more and more painful to use. As the number of sites gets larger, the user experience gets worse."
General Magic's Tabriz AgentWare and AgentTools allow you to create a server that hosts "intelligent" agents. Users can log on to change the profile of their agents and can be notified of results by e-mail, fax, pager, or by just checking the Web page used to configure the agent.
"The fundamental benefit is you can make all of this information accessible in a far more graceful manner to the end user," Kelsch said. "If you are a corporate MIS manager and are responsible for a large intranet with financial information, accounting information, or corporate information, and users want to extract information from it, you can create an agent, allow the user to specify information of interest to him and provide a whole new level of service to the user. Another side benefit is that you can reduce traffic on the intranet." (Editor's note: As this story went to press, Kelsch left General Magic.)
A number of companies have announced tools or services they expect will work with Tabriz on the Internet. Driftwood Systems, based in Bothell, WA, is creating a framework and application code, called puma, that will allow service providers to run personalized delivery services for their users.
"Right now if you use Telescript, said Steve Knox, president of Driftwood Systems, "you have to build your application from scratch. We are creating a set of classes for Web providers to put together a personalized delivery service."
To use the first version of puma you still need to learn Telescript, but subsequent versions will be easier to customize and will support Java on the client, Knox said. This will enable an information push to the client without having to wait for users to pull down a page. Knox said puma should be available for sale by the first quarter of 1997.
Agents don't have to be smart
Traditionally, agents have been though of as Net cruisers looking for something based on internal rules. However, these kinds of agents will always be limited by the amount of horsepower and artificial intelligence (AI) on the agent server. Another school of thought now recognizes that humans can improve an agents performance by offering feedback to the agent's findings.
Sapient Health Networks (SHN) is building a personalized Web information service around the Tabriz environment. It is designed as a medical information support system for people with chronic illnesses. The first disease to be supported will be fibromyalgia in October, followed by prostate and breast cancer by the end of the year.
James Kean, president of SHN, said that when patients are diagnosed with chronic diseases, they are lucky to get a 10 minute explanation from their doctor along with a skimpy brochure. But these diseases obviously have a profound impact on their lives, and the afflicted want to gather as much relevant information as possible. "Most patients do not have the time, expertise, or energy to sift through the vast amounts of information now on the Web. We are able to offer a service that acts much like a patient's own staff of researchers but exists only in cyberspace as intelligent agents."
When users first log onto the system, they answer a few questions which enable the system to generate a profile of them. Then over time, they rate all of the documents that are useful to them, and this information is fed back into the system to help others with similar profiles.
For now the operation is fairly people-intensive as Sapient begins to build a knowledge base of crucial information. A dozen full-time researchers work on each disease to filter out information, but Kean predicts this number will drop to three or four people per disease in 18 months, once the process becomes more automated.
Using agents to statistically match subjects of interest with particular people isn't a new concept. Max Metral, chief technology officer at Firefly Network Inc. said, "We started working on agents at MIT's Media Lab about five years ago. When we realized that these agents had to learn from each user individually we wondered why should they not leverage other agents to become more useful more quickly. At the same time we started worrying about domains like music where human feedback is the most useful thing to go on. The real problem with AI and agents is that people were trying to make them super intelligent. We said we will just leverage what works, which is people."
The group ran "Ringo" in the lab in 1993 as a way of automating the word-of-mouth process for music selection. Ringo was an e-mail-based program for matching musical interests. Users sent in a form with a list of music they liked, and Ringo mailed back a response. That was followed by "Homer," a Web-based music-matching system similar to Ringo except that users could submit queries and get results on the Web. Both programs ran on a single Alpha workstation, and neither was very scalable. The group's projects have since evolved into Firefly, an advertising-driven site that helps people find movies and music based on the selections of others with similar tastes.
Firefly is also making this technology available to other content providers for use on their servers. Yahoo, Ziff-Davis, Reuters, and Rolling Stone have all announced plans to run Firefly on their services. To get the service up and running, webmasters add a few HTML tags to their pages and run the Firefly software on a Unix or NT server.
"Our goal is to provide this functionality across many sites as a tool for providing personal information and targeted content to end users because its value comes in when you can go anywhere and you only have to teach one agent," Metral said.
Firefly makes all of its money from advertising, licensing, and selling data. But it does not sell contact data to marketers for fear of scaring away users or encouraging bogus name entry. In fact the company has even brought in the accounting firm Coopers and Lybrand as a public relations stunt to audit its information dissemination policy and assure users their personal information is not leaked to marketers.
"It is our core mission to say we will never sell your exact data. There are three kinds of data: demographic, preference, and contact. We don't have to give people contact data for it to be relevant. They can get lots of information out of demographic and preference data," Metral said.
Java promises open agents
From a webmaster's perspective, a limitation of using tools like Tabriz are their proprietary natures. You can only deploy them if you buy a server from one company, General Magic. Ideally, you should be able to write the agent once and then run it on any Web server. Java promises this, but standards still have to be worked out for moving agents over the Internet.
One company expanding on both Telescript and Java-based approaches is Kinetoscope. Its work with the Java Agent Workshop will allow users to create a variety of agents for different tasks. The preliminary trial version was due in late September. One of the first kinds of agents under development will be able to track stocks and report back by way of pager, Web page, e-mail, or fax. Another will page patients when it's time for them to take medication. Another agent will monitor your Web site and page you if it goes down. The environment is extensible, so once you have developed a way of tracking the information you want, you can just plug in the module for delivering the information in the way you need, whether it is fax, page, or voicemail.
"There are a number of environments for mobile agent code," said Damien Miller, director of engineering at Kinetoscope. "The features we are developing are not tied to any of those. We have had a fruitful relationship with General Magic, but there are a lot of business decisions that go into what tools to use. For instance, I anticipate a lot more support for JDBC [JavaSoft's SQL database access interface] database connectivity in the short term than for database connectivity to Telescript. There is a larger community working on Java application connectivity to databases and that translates into more flexibility in the long run."
Miller said, "Every good tool I have seen beside Telescript is executed in Java. The problem with Java is that it has a much more chaotic feel. The remote method invocation from JavaSoft allows mobile programs to move from computer to computer, but that API has not been cast in stone yet. JavaSoft has put an API out for comments which is subject to change so if you develop software to it, you are probably going to be hung out to dry because it will not support all of the features in the final release. Telescript is well defined, but Java is more widely accepted, and in the long run it will be a longer-term solution. I would guess we would move toward Java as the technology improves."
Aglet in, aglet out
One of the current limitations of Java for running agents is that it's only designed to move applets over the network, but not their data or their state of execution. IBM is working on combining Java applet technology with agents into what it calls "aglets." Aglets can move from site to site as they run since they carry their own data and execution state with them.
IBM is championing this as a standard and has already submitted a proposal to the Object Management Group. A preliminary Aglet Workbench which provides a visual environment for developing mobile applications is available at its Web site. It has a data access component that lets you access any ODBC-compliant database. In addition, IBM planned to have a JDBC driver online by the end of September.
Dan Chang, senior programmer at IBM, sees three areas where aglets could prove useful. They could be used to look through data at different sites to find something important for the sender. Mobile users could use them to find resources and complete tasks for them while they are off-line. They could also be used in network management to notify appropriate people when something breaks down.
Chang said IBM chose to work with Java because "the Java virtual machine will be available on any platform, and the Java byte code can be deployed anywhere. If you use Telescript, then the problem is that people have to buy from General Magic, and you have to make sure it is available on all platforms."
Chang sees two challenges in the deployment of aglets. First they have to provide more value than existing programming technologies. For example, search engines are widely available, so an aglet-based search engine would have to add some kind of unique value, or make it vastly easier to program. "A lot of time you have an application that can be solved with mobile agents that can also be solved with other more common stuff, and the added value of moving to mobile agents will not be realized," Chang said.
The other challenge lies in security. As soon as you begin to allow mobile agents to run around your network collecting data, you lose an element of control since hackers will have one more door into your enterprise. "People will more likely use it for intranet mission-critical applications rather than the Internet because it is easier to ensure security," Chang said.
Agents are not going away
Agent technology promises to simplify our lives with its ability to customize information for individual needs. But at the moment, the technology is only in its early stages and it will raise new legal questions as it becomes more sophisticated. What will be the legal ramifications of agents that make deals on our behalf. Will you have to keep the car it bought for you over the Net, even if you don't really want it?
Still, it is only a matter of time before agents are used to complete boring tasks. Wouldn't it be great if you were on a Web page with a long form that had to be filled out before you got some coveted software, and you just pressed a button to automatically have your agent do it for you? Or better yet, you could create an agent that goes to Web sites that pay you to look at ads, like CyberGold, and let it rake in the gold while you sleep.