A multi-agent system is a network of autonomous computational entities — agents — that share an environment and pursue their own objectives while interacting with one another. The discipline grew out of distributed artificial intelligence in the late 1980s and matured through the 1990s into a field with its own theory, architectures, and engineering practice. The Agentcities initiative was, in effect, an attempt to take the abstractions of multi-agent systems research and deploy them on the open internet at a scale and degree of heterogeneity that laboratory experiments had never approached. To understand what Agentcities was trying to demonstrate, it helps to understand the conceptual foundations on which it rested.
What Is a Multi-Agent System
At its simplest, a multi-agent system is a collection of agents situated in a common environment, where each agent perceives some portion of that environment through its sensors, decides what to do through its reasoning, and effects change through its actuators. Unlike a single monolithic program that solves a problem through one centralized control flow, a multi-agent system distributes problem solving across many autonomous agents that may cooperate, compete, or simply coexist. This distribution is not merely an implementation detail. It changes the character of the system fundamentally: there is no global controller, no single point that holds the complete state, and no guarantee that any one agent has an accurate or complete picture of the whole. Robustness, flexibility, and the capacity to scale across organizational boundaries all flow from this design, and so do its hardest problems.
Core Agent Properties
The canonical definition of what makes a piece of software an agent comes from Michael Wooldridge and Nicholas Jennings, whose 1995 survey identified four properties that have anchored the field ever since. Autonomy means the agent operates without direct external intervention, exercising control over its own actions and internal state. Reactivity means it perceives its environment and responds in a timely fashion to changes within it. Proactiveness means it does not merely react but takes the initiative, pursuing goals of its own rather than waiting to be told what to do. And social ability means it interacts with other agents — and sometimes with humans — through some form of agent communication language. An entity that exhibits all four is what Wooldridge and Jennings called a strong notion of agency, and it is precisely this combination that Agentcities required of the software populating its testbed.
BDI Architecture
The most influential internal architecture for a deliberative agent is the Beliefs-Desires-Intentions model, usually abbreviated BDI. Its philosophical grounding lies in Michael Bratman's 1987 account of human practical reasoning, which argued that intentions are not reducible to beliefs and desires but play a distinct functional role in committing an agent to a course of action. That insight was formalized computationally through influential work at SRI International and the University of Melbourne, producing a family of agent architectures and programming languages. In the BDI model, beliefs represent the agent's knowledge about the world, which may be incomplete or even incorrect; desires represent the motivational states or goals the agent would like to bring about; and intentions represent the desires the agent has committed to and is actively pursuing through an executing plan. The BDI control cycle runs continuously: the agent perceives, updates its beliefs, deliberates over its desires, selects an intention to commit to, executes the corresponding plan, and then repeats. This architecture matters for interoperability because FIPA-ACL is itself grounded in a BDI-style semantics, giving each message a formal meaning expressed in terms of the sender's and receiver's mental states rather than merely its syntactic form.
Emergence and Collective Behavior
One of the defining and most studied phenomena in multi-agent systems is emergence — system-level behavior that is not present in, or readily predictable from, the behavior of any individual agent. The textbook example is ant colony optimization, in which simple agents following local rules about depositing and following pheromone trails collectively discover efficient paths that no single ant comprehends. Emergence can be a feature to be harnessed or a hazard to be guarded against. In a global, open testbed such as Agentcities, emergent behavior could be benign and even desirable — for instance, novel and unanticipated combinations of services arising as agents composed one another's capabilities. But it could equally be problematic, manifesting as cascading failures or message storms that propagated across the federation faster than any operator could intervene. Understanding when emergence is beneficial, how to encourage the productive kind, and how to contain the destructive kind was a live and unresolved research question throughout the network's lifetime.
Coordination and Negotiation
Because a multi-agent system has no central controller, the agents within it must coordinate their activities through interaction rather than command. A foundational mechanism is the Contract Net Protocol, in which an agent with a task to delegate broadcasts an announcement, interested agents submit bids, and the announcing agent awards the contract to the most suitable bidder. Auction mechanisms generalize this idea of competitive allocation: English auctions raise the price by ascending bids, Dutch auctions lower an offered price until someone accepts, and combinatorial auctions allow agents to bid on bundles of items whose value is interdependent. Beyond simple bidding, argumentation-based negotiation lets agents exchange not just offers but the reasons behind them, enabling more flexible agreements when preferences are complex or partially aligned. Social norms and explicit organizational structures provide a further layer of coordination, constraining behavior so that individual autonomy does not collapse into chaos. The Association for the Advancement of Artificial Intelligence has published extensive research on multi-agent coordination, and these mechanisms were far from academic for Agentcities: its working groups on Business Process Integration (BPI) and on Service Directory and Composition (SDC) grappled directly with how independently operated agents could coordinate to deliver composite services across organizational boundaries on a running network.