AI Agent概论：从大语言模型到自主智能体 / Introduction to AI Agents: From LLMs to Autonomous Agents

AI Agent概论：从大语言模型到自主智能体

1.1 引言：从语言模型到智能体的范式跃迁

The landscape of artificial intelligence has undergone a profound transformation in recent years, evolving from static, task-specific language models to dynamic, autonomous systems capable of reasoning, planning, and taking actions in complex environments. This paradigm shift, centered on the emergence of Large Language Model-based AI Agents, represents not merely an incremental improvement in capability but a fundamental reconceptualization of how intelligent systems can interact with the world. The journey from symbolic artificial intelligence through expert systems to the contemporary era of LLM-driven agents encapsulates decades of theoretical development, practical experimentation, and technological advancement, each phase contributing essential insights that have culminated in the sophisticated agent architectures we observe today.

The term "agent" in artificial intelligence carries substantial theoretical weight, drawing upon decades of research in distributed systems, cognitive science, and decision theory. At its core, an AI Agent can be understood as an autonomous entity that perceives its environment through sensory inputs, processes and reasons about that information using internal representations, formulates plans or strategies to achieve specified objectives, and executes actions that modify the environment or its own state. This conceptualization, while seemingly straightforward, encompasses a rich tapestry of design considerations, implementation challenges, and theoretical nuances that have occupied researchers and engineers since the earliest days of the field.

The integration of Large Language Models as the cognitive substrate for autonomous agents marks a watershed moment in this long evolution. Unlike their predecessors, which required explicit programming of rules, heuristics, and domain knowledge, LLM-based agents leverage the emergent capabilities that arise from training on massive corpora of human-generated text. These models exhibit remarkable abilities in natural language understanding, generation, reasoning, and few-shot learning that enable new categories of agentic behaviors previously considered beyond the reach of computational systems. The transformation is perhaps best understood by examining the historical trajectory that has led us to this point, tracing the conceptual and technical developments that have progressively expanded what artificial agents can perceive, reason about, and accomplish.

1.2 人工智能智能体的学术谱系与技术演进

1.2.1 符号人工智能与逻辑推理系统

The intellectual foundations of AI agents can be traced to the symbolic artificial intelligence paradigm that dominated the field from its inception in the mid-twentieth century through much of the 1980s. Symbolic AI, also known as "good old-fashioned AI" (GOFAI) in Rodney Brooks' memorable phrase, operated on the fundamental assumption that intelligence could be achieved through the manipulation of explicit, human-readable symbols according to formal rules. This approach gave rise to sophisticated systems for logical reasoning, automatic theorem proving, and the formal representation of knowledge, laying the groundwork for agent-like behaviors in computational form.

The development of logic-based reasoning systems during this period produced several influential frameworks that remain relevant to contemporary agent design. Resolution theorem provers demonstrated that complex logical inference could be automated, establishing principles of systematic search and deduction that continue to inform reasoning architectures. The Situation Calculus, introduced by John McCarthy in the 1960s, provided a formal framework for representing states of the world and actions that change those states, enabling agents to reason about the consequences of their actions before executing them. Similarly, the Event Calculus and various modal logics offered alternative formalisms for temporal reasoning, each with distinct expressiveness characteristics and computational properties.

Expert systems, which emerged as a practical application of symbolic AI principles in the 1970s and 1980s, represented one of the first attempts to deploy agent-like architectures in real-world settings. These systems encoded domain-specific knowledge in the form of explicit rules or frames, typically organized around an inference engine that applied logical deduction to draw conclusions or recommend actions. MYCIN, developed at Stanford in the early 1970s for medical diagnosis, exemplified this approach, employing a rule-based reasoning system that could infer likely bacterial infections from patient symptoms and laboratory results. The success of such systems demonstrated the potential value of automated reasoning but also revealed fundamental limitations that would drive subsequent research.

The brittleness of purely symbolic systems became increasingly apparent as researchers attempted to scale these approaches beyond narrow domains. The knowledge acquisition bottleneck—the difficulty of eliciting, formalizing, and encoding human expertise into machine-readable rules—proved to be a persistent challenge. Furthermore, symbolic systems struggled with the inherent uncertainty, ambiguity, and incompleteness characteristic of real-world information. These limitations would eventually motivate the shift toward statistical and neural approaches that have come to dominate contemporary artificial intelligence research.

1.2.2 行为主义智能体与反应型架构

The reactive architecture movement emerged in the late 1980s and 1990s as a direct response to the perceived limitations of symbolic AI, drawing inspiration from biological systems and emphasizing direct coupling between perception and action. Rodney Brooks' subsumption architecture, introduced in 1986, became the paradigmatic example of this approach, dispensing with centralized reasoning in favor of distributed, layered control systems where simple behaviors were composed to generate complex overall behavior. Brooks' influential argument that "the world is its own best model" challenged the assumption that intelligent behavior required explicit internal representations of environmental state.

The behavior-based robotics paradigm championed by researchers such as Ronald Arkin brought additional theoretical sophistication to reactive agent design. Arkin's Motor Schema theory proposed organizing behaviors as schema units that could be dynamically combined based on environmental conditions, enabling agents to flexibly respond to varying situations without extensive deliberation. This work drew connections to biological motor control and ecological psychology, enriching the conceptual foundations for agent architectures with insights from biological intelligence.

Reactive agents excel in environments where rapid response is essential and where the relevant environmental features can be directly perceived without complex inference. However, the limitations of purely reactive approaches became apparent when agents needed to operate in partially observable environments, pursue goals requiring extended sequences of actions, or coordinate with other agents across time and space. These challenges motivated hybrid architectures that combined reactive capabilities with more deliberative reasoning components, setting the stage for the sophisticated multi-layered agent designs that characterize contemporary systems.

1.2.3 信念-愿望-意图架构与理性智能体

The Belief-Desire-Intention (BDI) model, developed substantially by Michael Bratman, Anand Rao, and Michael Georgeff in the late 1980s and early 1990s, introduced a principled framework for understanding rational agency that has profoundly influenced agent system design. Drawing on Bratman's philosophical analysis of human practical reasoning, the BDI paradigm conceptualizes agents as maintaining explicit representations of their beliefs (information about the world), desires (states of affairs the