AI智能体概述：从概念到定义的演进

AI智能体概述：从概念到定义的演进

AI Agent Overview: The Evolution from Concept to Definition

当我们谈论AI智能体时，首先要回答一个看似简单却极其重要的问题：什么使得一个系统成为"智能体"而非仅仅是执行预定义逻辑的传统软件？这个问题不仅关乎术语的精确性，更直接影响我们如何设计、评估和部署这类系统。在过去几十年里，人工智能领域对"智能体"这一概念的理解经历了深刻的演变——从早期基于符号逻辑的理性智能体，到连接主义时代的感知-动作系统，再到如今大语言模型驱动的自主智能体，每一次范式转换都重新定义了我们对智能体的期待。本章将系统梳理这一演进历程，建立对AI智能体的统一认知框架，并详细解析现代智能体系统的四大核心支柱。

When we talk about AI agents, the first question we must answer seems deceptively simple yet profoundly important: what makes a system an "agent" rather than merely traditional software executing predefined logic? This question not only concerns the precision of terminology but also directly influences how we design, evaluate, and deploy such systems. Over the past few decades, the understanding of the "agent" concept in artificial intelligence has undergone profound evolution—from early rational agents based on symbolic logic, to perception-action systems of the connectionist era, and now to autonomous agents driven by large language models. Each paradigm shift has redefined our expectations of agents. This chapter systematically traces this evolutionary journey, establishes a unified cognitive framework for AI agents, and provides a detailed analysis of the four core pillars of modern agent systems.

---

AI智能体的历史演进：从符号主义到连接主义再到LLM时代

The Historical Evolution of AI Agents: From Symbolism to Connectionism to the LLM Era

理解现代AI智能体的本质，需要我们将视野放回到人工智能发展的历史长河中。智能体概念的演进并非孤立发生，而是与整个人工智能学科的范式转换紧密交织。要真正把握当代LLM驱动智能体的独特性，必须首先理解前两个时代智能体的设计哲学和能力边界。

To understand the essence of modern AI agents, we need to place our perspective within the long river of artificial intelligence development. The evolution of the agent concept did not occur in isolation but was closely intertwined with paradigm shifts throughout the AI discipline. To truly grasp the uniqueness of contemporary LLM-driven agents, we must first understand the design philosophy and capability boundaries of agents from the two preceding eras.

符号主义时代的理性智能体（1956-1990）

Rational Agents in the Symbolist Era (1956-1990)

人工智能的符号主义传统起源于1956年的达特茅斯会议，其核心假设是：智能行为可以通过符号操作实现，知识和推理可以被形式化为明确的规则和逻辑表达式。在这一范式下，"智能体"被理解为一种基于知识的系统，它维护一个内部的知识表示，通过逻辑推理引擎处理输入，并产生符合预定规则的输出。

The symbolic tradition of artificial intelligence originated from the 1956 Dartmouth Conference, with its core assumption being that intelligent behavior could be achieved through symbol manipulation, and that knowledge and reasoning could be formalized as explicit rules and logical expressions. Under this paradigm, an "agent" was understood as a knowledge-based system that maintained an internal knowledge representation, processed inputs through a logical inference engine, and produced outputs conforming to predetermined rules.

这种理性智能体的典型代表是专家系统（Expert Systems）。以MYCIN系统为例——这是一个1970年代由斯坦福大学开发的医学诊断专家系统——它通过维护一个包含数百条规则的知识库来实现感染疾病诊断。系统的运作机制是：接收患者症状和检验结果作为输入，根据规则库进行前向或后向链式推理，最终输出诊断结论和建议治疗方案。

The archetypal example of this rational agent is the Expert System. Take the MYCIN system as an instance—a medical diagnosis expert system developed at Stanford University in the 1970s. It implemented infectious disease diagnosis by maintaining a knowledge base containing hundreds of rules. The system's operational mechanism was: receiving patient symptoms and test results as input, performing forward or backward chain reasoning based on the rule base, and finally outputting diagnostic conclusions and recommended treatment plans.

这种设计哲学的特点在于透明性和可解释性。规则和知识都以明确的形式化语言表示，人类专家可以检查、验证甚至直接修改系统的决策逻辑。然而，这也带来了固有的局限性：知识的获取依赖于人工编纂（knowledge engineering），系统的能力严格受限于预定义的规则集合，无法处理超出知识库范围的异常情况，更谈不上从经验中学习和自我改进。

The defining characteristics of this design philosophy are transparency and interpretability. Rules and knowledge were all represented in explicit formal languages, allowing human experts to inspect, verify, or even directly modify the system's decision logic. However, this also brought inherent limitations: knowledge acquisition relied on manual encoding (knowledge engineering), the system's capabilities were strictly bounded by the predefined rule set, it could not handle exceptions beyond the knowledge base's scope, and it had no capacity for learning from experience or self-improvement.

连接主义时代的感知-动作智能体（1990-2017）

Perception-Action Agents in the Connectionist Era (1990-2017)

1990年代以后，随着计算能力的提升和机器学习技术的成熟，人工智能进入了连接主义时代。这一时期的核心转变是从手工编码知识转向从数据中自动学习。智能体的概念也随之演化：从依赖显式规则推理的理性系统，转变为能够从环境交互中学习和适应的主体。

Following the 1990s, with the advancement of computing power and the maturation of machine learning technology, artificial intelligence entered the connectionist era. The central transformation of this period shifted from manually encoded knowledge to automatic learning from data. The concept of agents evolved correspondingly: from rational systems relying on explicit rule-based reasoning to entities capable of learning and adapting through environmental interaction.

智能体与环境交互的范式在Rodney Brooks的包容式行为架构（Subsumption Architecture）中得到经典体现。Brooks在1991年提出的这一架构摒弃了传统的中央推理模块，主张智能行为可以通过多个相对简单的行为层次叠加产生。在他的设计中，一个移动机器人可能包含"避开障碍物"、"向光源移动"、"随机漫游"等多个竞争性的行为模块，这些模块直接感知环境并产生动作响应，无需复杂的符号表示和逻辑推理。

The paradigm of agents interacting with environments found its classic expression in Rodney Brooks' Subsumption Architecture. This architecture, proposed by Brooks in 1991, abandoned the traditional central reasoning module and advocated that intelligent behavior could emerge from the superposition of multiple relatively simple behavior layers. In his design, a mobile robot might include competing behavior modules such as "avoid obstacles," "move toward light sources," and "wander randomly." These modules directly perceived the environment and produced action responses without requiring complex symbol representation or logical reasoning.

这一时期的另一个重要发展是强化学习（Reinforcement Learning）在智能体设计中的应用。强化学习智能体通过与环境的持续交互来学习最优策略，其目标是最大化累积奖励信号。与基于规则的专家系统不同，强化学习智能体能够发现人类设计者可能未曾预见的解决方案，并在动态环境中表现出适应性。DeepMind的AlphaGo——2016年击败世界围棋冠军——就是这一范式的杰出代表，它通过数百万局自我对弈学会了远超人类经验的策略。

Another significant development of this period was the application of Reinforcement Learning (RL) to agent design. Reinforcement learning agents learn optimal strategies through continuous interaction with the environment, with the goal of maximizing cumulative reward signals. Unlike rule-based expert systems, reinforcement learning agents can discover solutions that human designers might not have anticipated, demonstrating adaptability in dynamic environments. DeepMind's AlphaGo—which defeated the world Go champion in 2016—serves as an outstanding representative of this paradigm, having learned strategies far beyond human experience through millions of rounds of self-play.

然而，这一时代的智能体也面临显著局限。强化学习智能体通常针对特定任务进行训练，缺乏跨领域的泛化能力；感知-动作系统虽然反应敏捷，却难以进行复杂的推理和长期规划；更重要的是，这些系统都缺乏对语言、常识和复杂上下文的深层理解能力。

However, agents of this era also faced significant limitations. Reinforcement learning agents were typically trained for specific tasks, lacking cross-domain generalization capability; perception-action systems, while responsive, struggled with complex reasoning and long-term planning; most importantly, these systems lacked deep understanding of language, common sense, and complex contexts.

LLM时代的自主智能体（2017-至今）

Autonomous Agents in the LLM Era (2017-Present)

2017年，Transformer架构的论文《Attention Is All You Need》发表，标志着自然语言处理和人工智能的转折点。此后几年，大语言模型（Large Language Models, LLMs）的能力经历了惊人的增长：从GPT-2到GPT-4，从BERT到Claude，从闭源模型到开源生态——这些系统不仅能够生成流畅的文本，更展现出对语言的深层理解、复杂推理乃至初步的规划和工具使用能力。

In 2017, the publication of the Transformer architecture paper "Attention Is All You Need" marked a turning point in natural language processing and artificial intelligence. Over the following years, large language models (LLMs) experienced astonishing capability growth: from GPT-2 to GPT-4, from BERT to Claude, from closed-source models to open-source ecosystems—these systems not only generated fluent text but also demonstrated deep language understanding, complex reasoning, and even preliminary plannin