Foundations and Formal Models of Multi-Agent Systems

1.1 Why Multi-Agent Systems? The Limits of Single-Agent Abstractions

A software agent acting in isolation is no longer an adequate abstraction for the computational systems that modern society deploys. Consider, by way of orientation, a contemporary logistics network in which thousands of autonomous vehicles, warehouses, retail points, and customer devices each make decisions under uncertainty while their fates remain tightly coupled through shared resources, shared deadlines, and shared information channels. Treating such a system as a monolithic optimization problem is not merely inefficient; it is conceptually misaligned with the situation. The system is decentralized by physical necessity, it is partial-information by construction (no agent observes the full state), it is heterogeneous by design (different agents have different capabilities and objectives), and it must scale to populations that change continuously. These four properties—decentralization, partial information, heterogeneity, and scalability—are the empirical conditions that single-agent abstractions cannot capture, and they jointly motivate the multi-agent abstraction that this book develops.

The central question of this chapter, and of the discipline that it introduces, is therefore the following: *what constitutes a multi-agent system, and what formal foundations are required to analyze such systems rigorously?* The answer cannot be merely definitional. A multi-agent system (MAS) is not a software architecture in the conventional sense; it is a mathematical object whose behavior emerges from the interaction of multiple decision-makers, each equipped with its own information, preferences, and computational apparatus. To reason about such a system, one needs a vocabulary that distinguishes the agent from its environment, formal primitives for the mental attitudes that govern its choices, and analytical frameworks that connect individual rationality to collective behavior. This chapter supplies that scaffolding. Sections 1.2 and 1.3 establish the agent-level vocabulary. Section 1.4 surveys architectures. Section 1.5 compares the principal analytical frameworks. Section 1.6 works through the Contract Net Protocol as a canonical example. The chapter closes by previewing the structural arc of the book.

Before proceeding, a terminological clarification is in order. The literature uses the term "agent" inconsistently, and the inconsistency is not accidental: it reflects genuine disagreement about the right level of abstraction. This chapter adopts the convention, due originally to Wooldridge and Jennings (1995), of distinguishing *weak* and *strong* agency, and adds a third category, *situated agency*, that captures the embodiment dimension emphasized by robotics and sensor-network research. The taxonomy is not merely philosophical; it determines which formal tools are appropriate for which class of systems.

1.2 Agents, Environments, and the Notion of Agency

1.2.1 Weak Notions of Agency

The weak notion of agency, originating in the distributed artificial intelligence literature of the 1980s, requires only that an entity exhibit four properties: autonomy (it acts without direct external intervention), social ability (it interacts with other agents or humans), reactivity (it perceives and responds to its environment), and pro-activity (it exhibits goal-directed behavior rather than merely responding to stimuli). This definition is deliberately minimal. It does not commit the agent to having beliefs, desires, or any other folk-psychological attitude; it does not require learning; and it does not presuppose that the agent acts rationally in any economic sense. The weak notion is broad enough to subsume thermostats, packet routers, market-making algorithms, and human decision-makers within a single conceptual framework, which is its principal virtue.

The weakness of the weak notion is precisely its breadth. Because almost any reactive control system qualifies as a weak agent, the definition has limited discriminating power when one must decide whether a given system should be modeled as a multi-agent system or whether a simpler abstraction (e.g., a coupled set of differential equations, or a single optimization problem with distributed sub-solvers) would suffice. The discipline of multi-agent systems therefore requires operational criteria that go beyond the four properties above. We return to these criteria at the close of this section.

1.2.2 Strong Notions of Agency

The strong notion of agency, by contrast, endows agents with mental attitudes: beliefs, desires, intentions, and the like. This folk-psychological vocabulary, originally imported into AI by the seminal work of Cohen and Levesque (1990) and Rao and Georgeff (1991), is not a metaphysical commitment. It is a modeling commitment: representing an agent by its beliefs and goals allows the engineer to specify its behavior declaratively, to verify its conformance to design specifications using modal logic, and to predict its behavior under novel circumstances through the axioms of rationality that constrain the relationships among attitudes. The most influential formalization, the Belief-Desire-Intention (BDI) architecture, has its operational counterpart in systems such as PRS, dMARS, JACK, and AgentSpeak(L).

Three technical points deserve emphasis. First, BDI logics are not merely descriptive; they are *normative* in the sense that they encode which combinations of mental attitudes are rational. For example, the canonical Rao-Georgeff axioms require that intentions be consistent with beliefs (an agent cannot intend an action it believes impossible) and that an agent's intentions be filtered by its beliefs about means-ends relationships. Second, BDI architectures separate the deliberative process (deciding *what* to intend) from the means-ends reasoning (deciding *how* to achieve an intention), which permits real-time reactivity even when deliberation is incomplete. Third, the strong notion introduces a *semantic gap*: between the symbolic representation (e.g., a database of beliefs) and the operational reality (e.g., noisy sensor data), there is no canonical mapping. Resolving this gap is one of the central engineering problems of agent-based systems, and Chapter 4 takes it up in detail.

1.2.3 Situated Agency

The situated notion of agency foregrounds embodiment. A situated agent is embedded in a physical or virtual environment whose dynamics it cannot fully control, whose state it can only partially observe, and whose actions it must execute under resource constraints (time, energy, communication bandwidth). The situated notion is dominant in multi-robot systems, sensor networks, and embodied AI, and it stresses properties that the weak and strong notions treat as secondary: localization, uncertainty quantification, communication as a costly resource, and robustness to environmental perturbations. From the situated perspective, the formal apparatus of BDI logic is at best a local description of an agent's higher-level reasoning; the global behavior of the system is governed by physical laws, communication protocols, and stochastic processes that the symbolic abstraction abstracts away.

The three notions are not mutually exclusive. A situated agent can be a strong agent (it has beliefs, desires, and intentions, in addition to a physical embodiment); a strong agent is typically weak (it satisfies the minimal four properties); and weak agency is the common foundation. The choice of which notion to adopt is a modeling decision driven by the analytical questions one wishes to pose. If the question is "does this robot satisfy its mission specification under bounded sensor noise?", situated agency suffices. If the question is "does this team of agents have a coherent joint plan?", strong agency is required. If the question is merely "how does the aggregate behavior of these interacting entities scale with pop