Introduction

We envisage the evolution of the Internet into a free-market information economy in which billions of software agents exchange a rich variety of information goods and services with humans and amongst themselves [Chavez and Maes, 1996, Eriksson et al., 1996, Tsvetovatyy et al., 1997, White, 1996]. This will inevitably occur as agents assume an ever more pervasive and responsible role in electronic commerce. Even more fundamentally, the proven ability of a free-market economy to adjudicate and satisfy the conflicting needs of billions of human agents recommends it as a decentralized organizational principle for billions of software agents as well [Miller and Drexler, 1988]. However, given that software agents can make decisions several orders of magnitude faster than humans, and are vastly less flexible and complex, it is quite conceivable that an agent economy would behave in ways that are entirely unfamiliar. It is thus legitimate to ask whether a free-market information economy is inherently capable of facilitating the interactions of billions of software agents; and if so, what are the minimal requirements on the infrastructure of such an economy and on the agents that populate it.

An unequivocal answer cannot be found in the literature. Previous research suggests that large systems of interacting, self-motivated software agents can be susceptible to the emergence of wild, unpredictable, disastrous collective behavior [Kephart et al., 1989, Kephart et al., 1990]. On the other hand, a large body of work on market mechanisms in distributed multi-agent environments suggests that efficient resource allocation or other desirable global properties may emerge from the collective interactions of individual agents [Kurose et al., 1985, Huberman, 1988, Clearwater, 1995].

Much of the latter work falls under the rubric of ``market-based control'', in which economic transactions are used to bring about some predefined, desired end [Birmingham et al., 1996, Wellman, 1993, Stonebraker and others, 1994, Clearwater, 1995]. Agents may be designed to cooperate [Huberman et al., 1996] or to compete [Hogg and Huberman, 1991], but so long as the aggregate evolves toward a globally defined optimum, the system as a whole is deemed successful. But in an open system like the Web, there is no global purpose being served by the collective of agents; in a sense, there is no collective. Agents' goals may be harmonious, conflicting, or unrelated, as the case may be [Rosenschein and Zlotkin, 1994]. One cannot prescribe a universal medium of exchange, a universal ontology of goods and services, or a universal set of agent types or algorithms. Rather, these must emerge as the system evolves.

All of this motivates a general, wide-ranging study of economically motivated autonomous agents. [Kephart et al., 1998, Hanson and Kephart, 1998] In this paper, we focus on one uniquely ``economic'' property, the price of information, and investigate the consequences of different price-setting algorithms in a simple model of a multi-agent news filtering economy inspired by information dissemination services that can be found on the Internet today. In Section 2, we describe the details of the model. Section 3 delineates the system's state space and presents a baseline analysis of its dynamical behavior under an idealized price-setting algorithm, in the case of direct competition between two brokers offering a single type of information good. Some details of the calculations are deferred to an Appendix. Section 4 presents numerical results for a more complex situation in which three brokers may choose freely among three types of information good. We close by discussing the generality of our results and indicating some future directions.