AI Agent 概论：概念、起源与范式转变 | Introduction to AI Agents: Concepts, Origins, and Paradigm Shifts

AI Agent 概论：概念、起源与范式转变

核心问题

What fundamentally distinguishes an AI Agent from a traditional AI system, and why does this distinction matter for next-generation AI applications?

This question lies at the heart of understanding the current technological transformation. Traditional AI systems, despite their sophisticated capabilities in pattern recognition, classification, and prediction, operate fundamentally as sophisticated functions—receiving input, producing output, and terminating. They lack the capacity to maintain state across interactions, to plan sequences of actions toward long-term objectives, or to autonomously decide when and how to engage with their environment. The AI Agent paradigm represents a qualitative shift: from AI as a tool to be wielded, to AI as an entity that acts in the world. This distinction is not merely semantic—it carries profound implications for how we design, deploy, and reason about artificial intelligence systems. Understanding why this matters requires tracing both the conceptual foundations of agency and the technological developments that have made modern AI Agents possible.

智能体的本体论追问：从哲学概念到计算实现

1.1 智能体概念的哲学渊源

The concept of an "agent" predates computing by millennia, rooted in philosophy's investigation of causality, intentionality, and purposeful action. In philosophical discourse, an agent is fundamentally that which acts, or brings about change—a being with the capacity to initiate actions toward ends. Aristotle's distinction between potentiality and actuality touches on this: an agent is something that has the capacity to do something, to bring about states of affairs. This philosophical foundation emphasizes two critical attributes: the capacity for autonomous initiation of action, and the directedness toward goals that may not be immediately present.

The philosophical notion of agency introduces several concepts that remain central to computational agent design. First, intentionality—the "aboutness" of mental states, the capacity to represent and be directed toward objects and states of affairs in the world. Second, autonomy in the strong sense: the capacity of an agent to govern its own actions, not merely executing pre-programmed responses but determining its own course of action based on its representation of the situation. Third, rationality as a guiding principle: agents are typically understood as acting for reasons, following patterns of inference that connect beliefs, desires, and intentions to actions.

These philosophical foundations matter because they illuminate what we are trying to achieve when we design computational agents. When we speak of AI Agents, we are not merely creating sophisticated pattern matchers—we are attempting to create systems that can represent situations, form intentions, make plans, and execute actions in pursuit of goals. The philosophical tradition reminds us that agency is not a single property but a cluster of interrelated capacities that must be understood and implemented holistically.

1.2 计算智能体的诞生与演变

The computational notion of an agent emerged in the 1980s and 1990s within the artificial intelligence community, particularly in the context of distributed AI and autonomous systems research. Franklin and Graesser's 1997 typology of autonomous agents identified a crucial distinction: agents versus software agents. True agents, they argued, perceive their environment through sensors, act upon the environment through actuators, and operate autonomously—that is, without direct intervention by humans or other agents. Software agents, by contrast, are merely agents metaphorically, acting "on behalf of" users in a delegative sense.

This distinction remains conceptually useful today, though the boundaries have become more nuanced with modern developments. The early computational agents were predominantly symbolic systems, built on the principles of classical AI: knowledge representation, logical reasoning, and planning algorithms. These systems implemented some aspects of agency—the ability to represent the world, reason about actions, and plan sequences of operations—but often struggled with the brittleness of symbolic representations and the frame problem of determining which aspects of the world are affected by actions.

The 1990s also saw the rise of reactive architectures, exemplified by Rodney Brooks' subsumption architecture, which rejected symbolic representation entirely in favor of direct coupling between perception and action. These systems demonstrated impressive robustness in dynamic environments but sacrificed the deliberative capabilities that define sophisticated agency. The debate between deliberative and reactive architectures reflected deeper questions about the nature of intelligence and the role of representation in agent systems.

1.3 智能体的核心属性界定

Building on this history, we can identify four defining attributes that distinguish genuine AI Agents from traditional AI systems:

自主性（Autonomy） refers to the capacity of an agent to operate without direct human intervention, controlling its own actions and internal states. Autonomy exists on a spectrum: at one end, agents that execute pre-defined procedures with minimal adaptation; at the other, agents capable of independently determining their goals and the means to achieve them. Modern AI Agents, particularly those powered by large language models, exhibit a qualitatively new form of autonomy—they can interpret instructions, identify relevant sub-tasks, and adapt their behavior based on accumulated experience, without requiring step-by-step human guidance.

反应性（Reactivity） is the capacity to perceive the environment and respond to changes in a timely fashion. Reactive agents monitor their environment and maintain an up-to-date model of its state, enabling them to notice and respond to opportunities and threats as they arise. This distinguishes reactive agents from batch processing systems that complete a fixed computation and terminate. Reactivity requires continuous perception, state maintenance, and the ability to trigger responses when conditions change.

社会能力（Social Ability） encompasses the capacity to interact with other agents and humans through some kind of communication protocol. Social agents can engage in negotiation, cooperation, competition, and coordination—adapting their behavior based on the actions and communications of others. In multi-agent systems, social ability enables complex emergent behaviors through local interactions, while in human-agent interaction, it enables more natural and effective collaboration.

主动性（Pro-activity） is the capacity to exhibit goal-directed behavior by taking initiative. Pro-active agents do not merely respond to their environment—they anticipate future states and opportunities, formulating and executing plans to achieve their objectives. This requires the ability to reason about the future, to formulate goals, and to engage in planning—capacities that go beyond simple stimulus-response mappings.

These four properties—autonomy, reactivity, social ability, and pro-activity—constitute what Wooldridge and Jennings termed the "weak notion of agency" in their foundational 1995 paper. Together, they define a conceptual space within which various instantiations of agents can be placed and compared. Traditional AI systems may exhibit some of these properties in limited ways, but it is the conjunction of all four that characterizes genuine AI Agents and distinguishes them from conventional computational systems.

范式转变的驱动力：从反应式到主动式系统

2.1 反应式系统的局限与突破

Reactive systems, in their computational sense, are those that respond directly to environmental stimuli without maintaining internal representations or engaging in deliberation. The quintessential example is a thermostat: it senses temperature, compares it to a setpoint, and turns heating or cooling on or off—without any model of why the temperature changed or what t