Sheila McIlraith (photo: supplied)
More than two decades before today’s surge of artificial intelligence (AI) agents, Sheila McIlraith and co-author Tran Cao Son published a seminal paper describing how intelligent agents could automatically combine services on the web to accomplish complex tasks.
“Adapting Golog for Composition of Semantic Web Services” was presented at the eighth International Conference on Principles of Knowledge Representation and Reasoning (KR) in 2002. The paper's lasting impact on the field was recognized with a Test of Time Award at the 23rd edition of the conference in Lisbon in July 2026.
In the early 2000s, the web was undergoing a transition.
“The web was being transformed from a place where documents and web-accessible programs — services — were predominantly utilized by humans to one that was increasingly being digested and utilized by agents, or programs that were operating over the internet,” recalls McIlraith, professor in the Department of Computer Science at the University of Toronto.
Researchers pursuing the emerging Semantic Web were exploring ways to make online information and services interpretable by computers, creating the possibility of software agents that could find, use and combine those services automatically.
Then a research scientist at Stanford University, McIlraith had previously published a high-impact paper introducing so-called “Semantic Web Services,” setting the stage for work that addressed the problem of how agents could determine how to combine services to accomplish tasks.
McIlraith and Son approached the problem by thinking of the web as a kind of robot: some services could act as sensors, gathering information such as the weather or the availability and price of flights, while others could act as actuators, performing actions such as buying a book.
Things became more complex when an agent needed to combine several of these services.
Consider planning a trip. Most people understand the basic steps involved — choosing transportation, booking accommodation and getting from place to place. The challenge is choosing among the many available options while accounting for an individual traveller’s preferences and constraints.
McIlraith and Son developed a way to give an intelligent agent that kind of flexible template, or “generic procedure,” and have it adapt the procedure to a user’s particular needs and the services available. Their approach adapted Golog, a logic-based programming language originally developed at the University of Toronto to serve as the generic procedure, combining it with novel search techniques that interleaved information-gathering with acting.
The research helped shape how Semantic Web services were represented for computers and influenced subsequent academic and industrial work on automating the composition of web services, some subsequently done by McIlraith's graduate students after joining U of T.
Today, McIlraith, who is also associate director and research lead at the Schwartz Reisman Institute for Technology & Society and a faculty member at the Vector Institute, sees another dimension of the paper's legacy as AI researchers revisit many of the same questions in developing agents that can use external tools.
“This work really anticipated what we now call ‘tool use’ in frontier AI models and agents some 25 years before its re-emergence,” she says.