Foundations of AI Agents: Concepts, Definitions, and the Modern Landscape

The concept of an intelligent agent—one that perceives its environment, reasons about its state, and takes autonomous action to achieve specific objectives—has captivated researchers, engineers, and thinkers for decades. Yet for most of the history of artificial intelligence, such entities remained largely theoretical constructs, confined to laboratory demonstrations and academic papers. The landscape has transformed dramatically in recent years. Today, AI agents operate as indispensable components of complex software systems, ranging from recommendation engines that curate our digital experiences to autonomous systems that navigate physical environments with increasing sophistication. Understanding these agents—their architectures, their capabilities, their limitations, and their profound implications—has become essential knowledge for anyone seeking to comprehend the trajectory of contemporary technology and its impact on society.

Consider the experience of using a modern AI assistant to plan a complex trip. The system does not merely respond to individual queries; rather, it maintains awareness of your stated preferences, retrieves relevant information from various sources, breaks down the overarching goal into sub-tasks, executes those tasks in sequence or parallel, handles unexpected complications, and presents you with a coherent plan that synthesizes flights, accommodations, and activities. This seamless orchestration of perception, reasoning, and action represents the practical manifestation of agent-based AI systems, and it hints at the rich theoretical foundations that underpin such capabilities. The purpose of this opening chapter is to establish a comprehensive conceptual foundation for the journey that follows, examining the historical evolution of AI agents, introducing essential terminology and frameworks, exploring the revolutionary role of large language models as cognitive engines, and surveying the contemporary landscape of agent development platforms.

1.1 Historical Evolution of AI Agents

The intellectual lineage of AI agents traces back to the very origins of artificial intelligence as a discipline. When Alan Turing proposed his influential framework for machine intelligence in the 1950s, he implicitly envisioned machines that could perceive and act upon their environment. The celebrated Turing test, however, focused primarily on linguistic behavior—the ability to convince a human evaluator of one's humanity through text-based conversation. This emphasis on conversational mimicry, while foundational, represented just one possible dimension of machine intelligence. The agent perspective emerged as researchers began to grapple with more embodied and situated forms of intelligent behavior.

1.1.1 The Reactive Revolution

The earliest serious attempts to create artificial agents drew inspiration from biological systems and emphasized direct coupling between perception and action. Rodney Brooks, working at the Massachusetts Institute of Technology in the 1980s, articulated a influential critique of traditional AI approaches that relied on symbolic representations and centralized reasoning. In a series of groundbreaking papers, Brooks argued that intelligence did not require complex internal representations of the world; instead, intelligent behavior could emerge from simple reactive mechanisms that mapped sensory inputs directly to motor outputs. This philosophy, which he termed "Nouvelle AI" or "behavior-based AI," produced remarkable results in robotic systems capable of navigating complex environments without the computational overhead of traditional planning systems.

The subsumption architecture, developed by Brooks and his students, exemplified this approach. Rather than building comprehensive world models and deliberative planning modules, subsumption architectures composed simple behaviors in layers, where higher layers could subsume—or override—the outputs of lower layers when appropriate. A robot might have a layer dedicated to obstacle avoidance, another to wall following, and yet another to goal seeking. The entire system operated in real-time, responding fluidly to environmental contingencies without the delays inherent in symbolic reasoning. This reactive paradigm found commercial success in applications ranging from robotic vacuum cleaners to industrial automation systems, demonstrating that sophisticated behavior need not require sophisticated cognition.

The success of reactive approaches illuminated a fundamental tension in agent design that persists to this day: the trade-off between deliberative sophistication and reactive responsiveness. Purely reactive systems excel at real-time performance and robustness to environmental uncertainty, but they struggle with tasks that require lookahead, planning, or coordination of multiple goals. Conversely, purely deliberative systems can construct complex plans but may prove too slow or brittle for dynamic real-world environments. The history of AI agent development can be understood, in large measure, as the search for appropriate balances along this spectrum.

1.1.2 The Classical Planning Era

While reactive systems demonstrated practical utility in constrained domains, the classical AI community pursued a different vision: agents that could construct explicit plans by reasoning about the structure of their environment and their available actions. The planning formalism that emerged in the 1970s and 1980s treated agents as reasoning systems that manipulated symbolic representations of states, goals, and actions. Given a description of the initial state, a specification of the goal condition, and a library of action schemas describing preconditions and effects, planning algorithms would search the space of possible action sequences to find a path from initial state to goal state.

This planning paradigm achieved significant theoretical advances and produced influential algorithms including STRIPS (Stanford Research Institute Problem Solver), planning graph-based approaches, and heuristic search planners. These systems demonstrated that remarkable problem-solving capabilities could emerge from the manipulation of symbolic structures, leading many researchers to believe that the integration of perception, planning, and execution represented a tractable engineering challenge. The Shakey robot project at SRI International, completed in the early 1970s, represented an early attempt to integrate perception, planning, and execution in a single system, combining computer vision, natural language processing, and path planning to accomplish manipulation tasks.

However, the classical planning paradigm revealed significant limitations when confronted with real-world complexity. The assumption of perfect, complete knowledge of the environment proved untenable in most practical applications. The frame problem—the difficulty of specifying which aspects of the world remain unchanged by actions—created representational challenges that resisted clean solutions. And the computational complexity of planning, while addressed by sophisticated heuristics, remained a fundamental constraint on the scale and complexity of problems amenable to exhaustive planning approaches.

1.1.3 The Probabilistic Revolution

The recognition of uncertainty as a fundamental feature of real-world environments drove the integration of probabilistic reasoning into agent architectures. The influential work of Stuart Russell and Peter Norvig, among others, articulated a framework for rational agents that explicitly modeled uncertainty in perception and action. Rather than seeking guaranteed plans that would achieve goals under all possible contingencies, probabilistic agents could reason about the likelihood of various outcomes and select actions that maximized expected utility given their current beliefs about the world.

This paradigm found mature expression in the Markov Decision Process (MDP) frame