Foundations of Intelligent Agency

Foundations of Intelligent Agency

What distinguishes an AI agent from a conventional software program? This question sits at the heart of a technological transformation that is reshaping how we conceptualize, develop, and deploy artificial intelligence systems. A traditional program follows predetermined instructions, executing the same logic repeatedly unless explicitly modified by a developer. An AI agent, by contrast, exhibits a qualitatively different kind of behavior—it perceives environmental stimuli, reasons about appropriate responses, and acts autonomously to achieve goals that may not have been explicitly anticipated by its creators. Understanding this distinction is not merely an academic exercise; it represents the conceptual foundation upon which all subsequent technical content in this field must be built.

The distinction between programs and agents carries profound implications for software architecture, development methodologies, and the very nature of human-computer interaction. When we speak of AI agents, we invoke a set of capabilities—autonomy, reactivity, proactivity, social ability—that extend far beyond the input-output mappings of traditional computing. An agent does not simply process data according to fixed rules; it interprets situations, formulates strategies, selects actions from a space of possibilities, and adapts its behavior based on accumulated experience. These capabilities emerge from the integration of multiple computational subsystems, each contributing essential functionality to the agent's overall intelligence.

This chapter establishes the conceptual groundwork necessary for understanding AI agent systems by tracing their historical development, examining their philosophical foundations, and introducing the fundamental architectural patterns that structure their design. We begin with an exploration of how the notion of agency in artificial intelligence has evolved from its origins in symbolic reasoning systems to the modern paradigm of large language model-based agents. This historical trajectory reveals both the continuity of core ideas across decades of research and the revolutionary nature of recent advances that have transformed agents from theoretical constructs into practical engineering systems.

The Historical Evolution of Intelligent Agents

From Symbolic Reasoning to Modern Systems

The conceptual roots of AI agents extend to the very founding of artificial intelligence as a discipline. In 1956, when John McCarthy and his colleagues convened the Dartmouth Conference and coined the term "artificial intelligence," they envisioned machines that could exhibit forms of reasoning traditionally associated with human cognition. The early vision was ambitious: automated reasoning, natural language processing, abstract concept formation, and problem-solving capabilities that would rival or exceed human performance across a wide range of domains.

The initial approach to achieving these goals centered on symbolic AI, also known as "good old-fashioned AI" or GOFAI (Good Old-Fashioned Artificial Intelligence). Symbolic AI systems represented knowledge explicitly using symbols—文字、概念、逻辑关系—that could be manipulated through formal reasoning rules. The architecture of these systems was fundamentally deliberative: they maintained explicit representations of the world and their goals, searched through spaces of possible actions, and selected courses of action based on logical inference.

The General Problem Solver (GPS), developed by Allen Newell and Herbert Simon in 1957, exemplifies this early approach. GPS was designed to solve any problem that could be formalized as a means-ends analysis, where the system compares its current state to a goal state and identifies operators capable of reducing the difference. While GPS proved limited in practice—it could solve toy problems but failed to scale to real-world complexity—it established foundational concepts that remain relevant: the representation of goals, the modeling of problem states, and the use of search to find solution paths.

The subsequent decades witnessed the development of increasingly sophisticated symbolic architectures. Expert systems emerged in the 1970s and 1980s as a practical application of symbolic AI, encoding domain-specific knowledge in the form of rules that could be applied to reach conclusions or recommend actions. Systems like MYCIN (for medical diagnosis) and DENDRAL (for chemical analysis) demonstrated that symbolic representations could capture meaningful expertise and provide useful recommendations in specialized domains. These systems embodied a form of agency—they perceived inputs (symptoms, evidence), reasoned about their significance (applying rules of inference), and acted (producing diagnoses or recommendations)—though their autonomy was constrained by the narrow boundaries of their encoded knowledge.

However, symbolic AI faced fundamental limitations that became increasingly apparent over time. The frame problem, identified by philosophers and AI researchers, highlighted the difficulty of specifying which aspects of the world are relevant to a given reasoning task without exhaustive enumeration. The symbolic grounding problem questioned how symbols could acquire meaning through purely formal manipulations, without connection to embodied experience or environmental interaction. And the brittleness of symbolic systems—their tendency to fail catastrophically when encountering situations outside their explicit knowledge base—suggested that pure deliberation was insufficient for robust intelligence.

The Reactive Revolution

In response to these limitations, a new paradigm emerged in the late 1980s and 1990s: reactive AI. Pioneered by researchers such as Rodney Brooks, who famously argued for "intelligence without representation," reactive systems emphasized direct coupling between perception and action, without elaborate internal representations or symbolic reasoning. Brooks' subsumption architecture organized behavior into layered levels, where simpler reactive behaviors (avoid obstacles, move toward light) could be superseded by more complex ones (explore, navigate) without centralized control.

The reactive approach drew inspiration from biological systems, particularly insects, which exhibit sophisticated behavior despite limited neural complexity. The insight was that much intelligent behavior emerges from the interaction between an agent and its environment, rather than from the manipulation of abstract symbols. Situated agents—systems that exist within and act upon a physical or virtual environment—could exploit environmental regularities and affordances rather than constructing complete world models.

This period also saw the rise of behavior-based robotics, where robots were designed as collections of relatively simple behaviors that competed or cooperated to produce overall action. The subsumption architecture and its descendants represented a fundamentally different conception of agency: one where intelligence was distributed across perception-action couplings rather than concentrated in a central reasoning engine. Agents following this paradigm were reactive in the truest sense, responding immediately to environmental stimuli without intervening deliberation.

The Cognitive Agent Renaissance

The pendulum of AI research did not settle permanently on pure reactivity. The 1990s and early 2000s witnessed a synthesis of reactive and deliberative approaches, recognizing that successful agents often require both rapid reactive responses and slower deliberative planning. This synthesis gave rise to the cognitive agent paradigm, which incorporated symbolic representations and reasoning while maintaining the agent's situatedness and embodiment.

The Belief-Desire-Intention (BDI) model, developed by Michael Bratman and adapted for artificial agents by researchers includi