X

"Intelligent agent" technology staging a comeback

Companies from start-ups to heavyweights are still investing heavily in agents, and the results may yet change the way people and computers relate to one another.

Paul Festa Staff Writer, CNET News.com
Paul Festa
covers browser development and Web standards.
Paul Festa
9 min read
Hyped only a couple of years ago as the next big thing in store for the Internet, intelligent agents have largely flopped in the marketplace as users balked at turning over everyday tasks and negotiations to software applications.

But companies from start-ups to heavyweights are still investing heavily in agents, and the results may yet change the way people and computers relate to one another.

The new generation of products goes beyond offering automated shopping to knit together increasingly complex personal communications options, from email to instant messaging and telephones.

Developers including General Magic, Microsoft, and Lucent are heavily backing voice interface research, for example, which is yielding a new class of agent technology that some analysts predict will find a receptive commercial audience.

Hewlett-Packard, meanwhile, is using agent technology to allow databases to share information more freely.

"Whether by voice...or by typing, I think agents are going to be helping more people use the Internet," said Ted Kunzog, analyst and editor for Internet Stock News. "Within a year or two you will not need to use a computer to use the Internet. Instead people will be speaking into a phone or a wristwatch, and some little software product will bring back what they're looking for. That's the promise of agents."

Consumer confusion about agents is no fluke: The category encompasses such a wide range of products and technologies that even developers disagree over what the term means.

Artificial intelligence academics despair over arriving at a universally agreed-upon definition, but among software sellers the term generally is used to denote software that automates certain computing functions and exercises some judgment on the user's behalf.

Another agent-like technology lets computers engage users in conversations. Called "chatterbots," this kind of agent is best exemplified by Artificial Life and Neuromedia, according to Kunzog. Another well-known conversational tool on the Web is Ask Jeeves.

Intelligent shopping
E-commerce agents have attracted much of the attention recently, a typical application being software that independently combs the Web for the best price on consumer goods or monitors auctions and notifies users when they are outbid.

The trouble with these types of applications, however, is that there's little business incentive for an e-commerce site to direct its users to the cheapest alternative on the Web or to canvass all available competitors.

Sabeer Bhatia is one e-commerce entrepreneur who takes a dim view of price comparison agents. Bhatia, who founded Hotmail and sold it to Microsoft for $400 million, is now at work on an e-commerce start-up called Arzoo that will integrate various tools and tasks for Web purchasing.

"The business model gets screwed up," Bhatia said. "There's an inherent conflict. If one of your partners is going to be thrown in the same mix as everyone else, there's no reason for you to get paid any money. We have made a conscious decision not to use this type of technology because there's no viable revenue model."

Another famed e-commerce application classified under agent technology is the practice of collaborative filtering, in which a user's purchase is compared to the buying habits of other consumers, and recommendations are made based on what others bought. Collaborative filtering has met with some privacy concerns but has enjoyed much broader market acceptance, with implementation at e-commerce heavyweight sites such as Amazon.com.

Much of the agent technology under development is directly commerce related.

"Originally the idea behind agents was to have a piece of software that knew your user ID and password that could do things you normally would do yourself," said Danny Lange, chief technologist for software developer General Magic. "The idea was that agents could talk to each other. This hasn't taken off."

The reason, according to Lange and others, is that consumers were uncomfortable with the idea that computers were talking to computers. General Magic's solution is to have computers report back from their conversations with each other and bring the human user in for the decision-making part at the crucial time.

Like other companies, General Magic crafted its new agent's strategy from the rubble of its initial efforts.

"General Magic years ago developed technology with which agents talked to other agents, and it was clearly a technology that was too early," Lange said. "Even today that tech is not deployed anywhere. We have tried to take the best ideas from that and get it into a more practical context."

Initially developed in 1994, General Magic's technology was called Telescript. It has found new life in the company's Kenya technology, which lets users request action by voice command over the phone, telling the agent to purchase a new ticket if a flight is canceled, for instance, or upping an auction bid.

General Magic has posted an audio demonstration of the technology as it's used with General Magic's Portico Virtual Assistant, a voice recognition product for messaging. General Magic said it will add Kenya technology to Portico this year and provide a software development kit with which others can implement it.

The language of agents
Kenya is based on Extensible Markup Language (XML), a protocol for the creation of industry- or task-specific markup languages. XML makes it easier for Web authors to make their documents more clearly and accurately readable by other computers. XML, it turns out, is ideal for the kind of work intelligent agents do. Kenya, for example, uses XML to define rules that agents follow in carrying out their tasks.

Another XML-based agent technology coming down the pipeline is Hewlett-Packard's E-speak initiative. E-speak is an application for building Web-based networks of interconnected services.

For example, a financial planning service built with E-speak could communicate with a mortgage or a banking service to automatically update the user's balances and fees and recalculate his or her retirement countdown based on changes in various accounts. Services on the E-speak platform can decide how to communicate vital information to the user based on his or her platform, choosing to send the news via email, phone, or instant message.

Rajiv Gupta, general manager of HP's E-speak initiative, agreed with Bhatia that a weakness of agents has been their tenuous connection thus far to money-making schemes.

"The problems agents had earlier came about because agents were primarily technology with no associated business model," Gupta said. "And there were not strong enough notions of security, auditing, and billing, and those are the kinds of things you require for creating any business that will function."

E-speak has signed on a number of businesses to try out the technology, including a number of wireless device providers such as Motorola, Nokia, and Ericsson. HP will release the beta, or trial, version of E-Speak, version 2.1, at its E-Services developers' conference in San Jose next month.

Microsoft's vision
Microsoft has long been interested in intelligent agents. At the height of the craze in 1997 it debuted its Agents technology for creating animated assistants on the desktop that use back-end reasoning services. But that's only a preview of the kinds of things the company is cooking up in its lavishly funded research wing.

A fundamental goal of Microsoft's research efforts with agents is to let computers and their users converse more fluently with each other. One of the prototypes being constructed within Microsoft's "conversational architectures project" is called the Bayesian Receptionist, a personal assistant named for the 18th-century English Presbyterian minister and part-time mathematician Thomas Bayes, whose work gave rise to the theory of "Bayesian logic," concerned with reasoning under uncertainty.

"Agents are often uncertain about what the user really wants; most agents technology hasn't taken uncertainty into account," said Eric Horvitz, senior researcher at Microsoft Research, where he manages the adaptive systems and interaction group. "Conversational architectures projects are focusing on the whole notion of what is it like when two people have a conversation and what kinds of skills that requires."

Microsoft Research is creating agents that can watch users and make decisions on their behalf, or in collaboration with them, Horvitz said. One system, dubbed "Priorities," observes the user's habits as he or she sorts through mail, noting what the user tends to read first or what gets deleted without being opened. As Priorities becomes more knowledgeable about the user's habits, it learns to score incoming mail on a scale of zero to 100 and present it to the user accordingly.

Priorities considerations include actual phrases in the message, the company organization chart, and whether the mail was sent only to the reader or to others as well.

When the user receives highly scored email, Priorities examines its user profile to determine the best way of alerting the user. Because the system can sense whether the user is at the computer, on the phone, or--by consulting the user's desktop calendar-- ven whether he or she is in a meeting, it can decide whether to call the user on a cell phone or leave the message in the email in-box.

"Priorities knows when to bother me," said Horvitz, who uses a prototype of the system. "A few weeks ago, I landed in Washington, D.C., and had not yet seen all this afternoon email. I turned on my cell phone and found I had five messages that the system had decided were important enough to send on the cell, and they were all critical."

Some users might find another function of the project, named Lookout, overzealous. Microsoft is working on getting the software to anticipate the user's goals and act on them. For instance, by reading email the user has opened, Lookout decides whether to pop up on the screen and ask whether the user wants to schedule an appointment. If the answer is yes, the receptionist system can bring up the desktop calendar and schedule the appointment, all using a speech-based interface.

Venturing further toward the science fiction realm of anthropoid computers, Microsoft Research aims to develop what it calls an "attentional user interface," or AUI. This sensory capability will let the computer know where the user is and where his or her attention is focused so the computer can make decisions on how best to stream information or services to him or her.

For example, if the computer can hear that the user is in the middle of a conversation, it could decide not to bother him or her with an email notification, or it could weigh the importance of the message against the risk of intruding on the conversation.

"The scarcest resource these days isn't computing power or bandwidth," Horvitz said. "It's human attention. If computers one day soon get a really true sense of what we're doing, they can offer rich services. For example, instead of a simple reminder of a meeting coming up, I might get a streaming description of the meeting. The computer could pull multiple applications based on what the user was doing at the time."

Lucent gets talking
Like General Magic and Microsoft, three-year-old AT&T spin-off Lucent Technologies is interested in what agents can do with a voice interface.

Lucent said it took a hard look at agents' failure in the marketplace and tailored its products to learn from early mistakes. High on the list of gaffes in the telecommunications industry was the complexity of the offerings.

"Service providers couldn't explain or sell them, and the appeal was narrow," said Dan Johnson, product marketing manager at Lucent. "These were products aimed at the road warrior."

So Lucent started looking at smaller, individual pieces of the integrated offerings and came up with narrower offerings with broader appeal. Call screening and a so-called privacy manager were the results.

Using Lucent's technology, Ameritech took the ability to screen a call and marketed it as a way to block telemarketers. Ameritech has been signing people up "far ahead of expectations," Johnson said. The privacy manager launched nine months ago.

Lucent doesn't plan to abandon the more complex, integrated offerings, but it will continue rolling out its offerings in bite sizes the market can digest, according to Johnson.

"The migration strategy is to go from simple to complex over time," Johnson said. "The virtual assistant, or e-secretary, hasn't been successful in the marketplace, so clearly there has to be some other approach to automation and efficiency.

"Right now there are too many features for the average consumer to comprehend and make a value justification. If there's a big price tag and they're not sure what to do with it, that's not a good selling equation."