Intelligent Agent
An intelligent agent is anything that perceives its environment through sensors and acts upon that environment through actuators, chosen so that its behavior achieves good outcomes. The framing was made the organizing idea of a whole discipline by Stuart Russell and Peter Norvig in the first edition of Artificial Intelligence: A Modern Approach (Prentice Hall, 1995), which recast AI as the study of agents rather than as a collection of techniques such as logic, search or neural networks. A human has eyes and hands, a robot has cameras and motors, a thermostat has a temperature sensor and a switch, and a software agent has keystrokes, files and network packets as its percepts and actions. The abstraction deliberately fits all of them.
The central technical move is to separate what an agent does from how it is built. The agent function maps any sequence of percepts to an action; the agent program is the concrete implementation that runs on some architecture and approximates that function. This lets a table-lookup chess player, a hand-written rule system and a trained network be compared as the same kind of object: a mapping from observations to behavior. It also exposes why a lookup table is hopeless in practice, since the number of possible percept sequences explodes, and so the real work of AI is finding compact programs that produce good behavior without enumerating the function.
Rationality is defined relative to a performance measure, not as omniscience or perfection. A rational agent, in the Russell and Norvig sense, selects for each percept sequence the action expected to maximize its performance measure, given the evidence of the percepts and whatever knowledge it has built in. The definition separates three things that popular talk blurs together: what the designer wants (the measure), what the agent knows (percepts plus prior knowledge), and what actually happens (which luck can spoil). An agent that crosses a street after looking both ways can be rational and still be hit by a falling object. Later editions stress that the measure should describe what the designer wants in the environment, not how the agent should behave, to avoid designing for the wrong target.
PEAS is the discipline of describing the task environment first, and it carries most of the practical value of the framework. Performance measure, Environment, Actuators and Sensors must be written down before an architecture is chosen. The textbook also classifies environments along dimensions that predict how hard the problem is: fully or partially observable, single or multi-agent, deterministic or stochastic, episodic or sequential, static or dynamic, discrete or continuous, known or unknown. Chess is fully observable and deterministic; driving is partially observable, stochastic, dynamic and continuous, which is why the two demand very different agent designs.
The agent designs form a ladder, and each rung exists because the previous one fails somewhere. A simple reflex agent reacts to the current percept only and fails whenever the percept is ambiguous. A model-based reflex agent adds internal state, such as remembering what is behind it while driving. A goal-based agent compares predicted futures with a goal, using search and planning, and a utility-based agent handles conflicting goals and uncertainty by weighing outcomes with a utility function, which is where decision theory enters. A learning agent adds a critic that scores behavior against the performance measure, a learning element that improves the rules, and a problem generator that proposes exploratory actions; the structure echoes older work on trial-and-error learning and underlies reinforcement learning.
A frequent misconception is that an agent must be a robot, be autonomous in a dramatic sense, or be conscious. The concept is intentionally thin: a thermostat qualifies as an agent, though a trivial one, and the word intelligent is carried entirely by the quality of the mapping from percepts to actions. Another pitfall is treating rationality as a claim about perfect behavior; real agents face limited computation, so the field also studies bounded rationality, a notion going back to Herbert Simon. A third is to confuse the formal agent of this textbook with the distinct line of multi-agent systems research in the 1990s, such as the survey by Michael Wooldridge and Nicholas Jennings, which stressed properties like autonomy, social ability and reactivity.
The vocabulary has returned in force with language-model agents. A modern LLM agent wraps a language model in a loop: it reads an observation, produces a thought or action as text, calls a tool such as a search engine, code interpreter or file editor, and feeds the result back as the next percept; the ReAct paper by Yao and colleagues (2022) popularized interleaving reasoning with actions. The PEAS discipline applies unchanged, and it explains typical failures: a vague performance measure invites reward hacking or runaway tool use, partial observability produces hallucinated state, and missing a model of the environment produces repeated mistakes. The decades-old agent framework still supplies the checklist for judging such systems.
Read intelligent agent as a design lens rather than a kind of machine: specify what counts as success, what the system can sense and do, and then ask how well its choices serve that measure in the environment it actually faces.