Various types of learning interface agents have been implemented [Dent et al.1992][Hermens &Schlimmer1993][Maes &Kozierok1993][Kozierok &Maes1993]. All of them are essentially designed to act in a stand-alone fashion or engage in restricted task specific communication with identical peers. Our agents not only come up to speed much faster, but also discover which of a large set of heterogeneous peers are useful consultants to know in particular domains.
Multi-Agent Systems research has concentrated on negotiation and cooperation strategies that are used by autonomous agents who must compete for scarce resources. Various formal protocols and frameworks have been proposed to model agent's intentions, domains and negotiation strategies [Rosenschein &Genesereth1985][Zlotkin &Rosenschein1993] based on various game-theoretic, logical, economic and speech-act models. While the analytic frameworks above are important, most are based on restrictive assumptions about the domain or the agents' capabilities and assume that the reason agents cooperate is because they need access to a shared resource or have multiple overlapping goals.
The Ontolingua tools [Gruber1993] and the work on the KQML Agent-Communication Language [Finin et al.1993] provide a way for agents using different ontologies to communicate effectively with each other and may be used to implement our collaborative architecture. Our research represents an actually implemented system in a real domain that shows the benefits of collaboration amongst agents.