多智能体系统导论：协作智能的工程视角

1.1 从单点智能到协同智能：问题驱动的引入

An intelligent system operating in isolation faces a structural ceiling. Whether it is a deep model serving predictions, a robotic arm executing pick-and-place, or a recommendation engine ranking items, the moment the environment becomes open (open-set, open-world, open-vocabulary), dynamic (non-stationary distribution, adversarial perturbation, concept drift), or distributed (geographically scattered data, latency-sensitive coordination, partial observability), the single-agent paradigm exhibits three recurring failure modes: *perceptual incompleteness*（感知不完备）、*epistemic monotonicity*（认知单调性）以及 *decision brittleness*（决策脆性）。These failure modes are not engineering accidents; they are consequences of a closed-world assumption baked into the architecture of a single agent.

This chapter begins with a deceptively simple question: why must we engineer intelligence collaboratively rather than scale it monolithically? The answer is not "because multi-agent systems are fashionable." The answer is that certain classes of problems — coordination under uncertainty, decentralized control, emergent fault tolerance — are provably intractable for a single decision-maker, and the engineering community has accumulated decades of evidence that decomposition plus coordination outperforms centralization plus scale in those classes. The remainder of this chapter formalizes what a multi-agent system (MAS) is, traces its lineage, contrasts it against the limits of single-agent intelligence, surveys its application domains, and demonstrates, through a concrete warehouse-scheduling case, how three heterogeneous agents produce emergent intelligent behavior that none of them possesses individually.

A working definition, precise enough for the rest of this book, is the following:

> A Multi-Agent System (MAS) is a set of autonomous, situated, and interactive computational entities — each perceiving its local environment, reasoning about partial observations, and acting to influence that environment — whose joint behavior, mediated by an explicit communication and coordination substrate, produces system-level capabilities that exceed the sum of their individual capabilities.

The phrase *"exceed the sum"* is not rhetorical; it is the operational definition of emergence (涌现性) that we will revisit in Section 1.7. For now, it suffices to note that emergence is the *design goal* of MAS engineering, and the rest of the chapter is, in essence, a controlled argument that this goal is both reachable and necessary.

---

1.2 智能体与多智能体系统的形式化定义及核心特征

1.2.1 智能体（Agent）的最小定义

At the risk of oversimplifying a literature with several competing definitions (Russell & Norvig, Wooldridge, Ferber), we adopt a minimal but operational characterization suitable for engineering practice. An agent is a tuple

$$\mathcal{A} = \langle \mathcal{S}, \mathcal{O}, \mathcal{A}, \mathcal{T}, \pi, \mathcal{U} \rangle$$

where:

• $\mathcal{S}$ is a **state space** the agent can inhabit;
• $\mathcal{O}$ is an **observation function** $\mathcal{O}: \mathcal{S} \times \mathcal{E} \rightarrow \mathcal{P}(\mathcal{S})$ mapping environment states and external events into the agent's perceptual frame, where $\mathcal{E}$ denotes the environment;
• $\mathcal{A}$ is an **action set** the agent can execute;
• $\mathcal{T}$ is a (possibly stochastic) **transition function** describing how the agent's actions modify local state;
• $\pi$ is a **policy** $\pi: \mathcal{P}(\mathcal{S}) \rightarrow \mathcal{P}(\mathcal{A})$ mapping belief states into action distributions;
• $\mathcal{U}$ is a **utility function** or reward model $\mathcal{U}: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ encoding the agent's objectives.

The tuple is not new — it is essentially a partially observable Markov decision process (POMDP) lifted to the agent level — but the framing matters. An agent is not merely a policy; it is a *situated* entity that persists over time, occupies state, and is held accountable for a utility. This is what distinguishes an agent from a stateless function or a one-shot inference call.

The five properties that distinguish an agent from an arbitrary program are widely recognized as autonomy（自主性）、reactivity（反应性）、pro-activeness（主动性）、social ability（社会性）以及 temporal continuity（时间持续性）. We will treat these not as buzzwords but as engineering contracts: each property corresponds to a design decision the engineer must explicitly commit to.

• *Autonomy* means the agent exercises control over its own internal state and action selection without external micromanagement. A web service that only responds to RPC calls is **not** an agent; a service that schedules its own polling, reasoning, and actuation is.
• *Reactivity* means the agent responds in a timely fashion to changes in its perceptual frame. The response is bounded by a deadline relevant to the environment.
• *Pro-activeness* means the agent exhibits goal-directed behavior — it does not merely react, it *pursues*.
• *Social ability* means the agent can exchange information with, and reason about, other agents through some explicit interaction protocol. This is the property that elevates an agent into a member of a MAS.
• *Temporal continuity* means the agent has a persistent identity and state across observation–action cycles, rather than being instantiated fresh per call.

In the contemporary Agentic AI discourse, these properties are sometimes blurred — large language model (LLM) agents are often described as agents even when they lack temporal continuity. We will return to this distinction in Chapter 2 when we model the internals of an LLM-backed agent; for now, we flag it as an open semantic question.

1.2.2 多智能体系统（MAS）的形式化定义

A multi-agent system is a tuple

$$\mathcal{M} = \langle \mathcal{A}_1, \dots, \mathcal{A}_n, \mathcal{E}_{\text{shared}}, \mathcal{C}, \mathcal{R}, \mathcal{G} \rangle$$

where:

• $\mathcal{A}_i$ for $i \in \{1, \dots, n\}$ are individual agents (each conforming to the tuple above);
• $\mathcal{E}_{\text{shared}}$ is the **shared environment** in which the agents are co-situated;
• $\mathcal{C}$ is a **communication substrate** — the channels, languages, and protocols through which agents exchange messages;
• $\mathcal{R}$ is a **coordination regime** — the explicit or implicit mechanism (auctions, contracts, voting, negotiation, emergence) by which joint decisions are reached;
• $\mathcal{G}$ is the **system-level objective** — the function the collective is engineered to optimize or satisfy.

Three structural properties of $\mathcal{M}$ deserve immediate emphasis.

First, partial observability is the default, not the exception. In a MAS, no single agent typically has access to the global state $\mathcal{S}_{\text{global}}$; each $\mathcal{A}_i$ perceives a slice $\mathcal{S}_i \subseteq \mathcal{S}_{\text{global}}$. Engineering under partial observability is one of the most consequential departures from classical single-agent AI: it forces explicit representations of *belief*, *uncertainty*, and *common knowledge*.

Second, the coordination regime $\mathcal{R}$ is a first-class design object, not an implementation detail. The choice between a centralized auctioneer and a peer-to-peer gossip protocol is not a "details later" question — it determines the system's communication complexity, fault tolerance, and even its attack surface. We will devote Chapter 5 to a systematic treatment of $\mathcal{R}$.

Third, the system-level objective $\mathcal{G}$ is generally not the sum of individual utilities. In a cooperative MAS, $\mathcal{G} \approx \sum_i \mathcal{U}_i$; in a competitive MAS, $\mathcal{G}$ may be Pareto-optimal over individual utilities but not summable; in a mixed MAS, $\mathcal{G}$ may even be in tension with $\mathcal{U}_i$. The relationship between $\mathcal{G}$ and $\{\mathcal{U}_i\}$ is the **alignment probl