An intelligent agent is a system that perceives its environment through sensors and acts on it through actuators to achieve goals. In classical…
Define the intelligent agent as a perceive-then-act system judged by a performance measure.
An intelligent agent is anything that perceives its environment through sensors and acts on it through actuators to reach a goal. At each step it maps what it has perceived to an action — a function from percepts to actions.
An agent is called rational when it picks the action expected to do best against its performance measure, given its knowledge. Rational does not mean perfect or all-knowing; it means making the best choice with the information available.
To describe an agent precisely, classical AI uses PEAS: the Performance measure (how success is scored), the Environment it operates in, its Actuators (how it acts), and its Sensors (how it perceives). For a self-driving car: performance is safe, fast, legal driving; the environment is roads and traffic; actuators are steering and brakes; sensors are cameras and radar.
The environment's properties — fully or partially observable, deterministic or stochastic, single- or multi-agent — shape how sophisticated the agent must be.
Learn the five agent types ordered by how much they model and optimize.
A simple reflex agent acts only on the current percept using condition-action rules: 'if condition then action.' It keeps no history and has no model of the world. It works only when the right action depends entirely on what it senses right now, and fails when the environment is partially observable.
A model-based reflex agent keeps an internal state — a model of how the world evolves and how its actions change it — so it can handle a partially observable environment. It uses that state to infer what it cannot directly see before applying its rules.
A goal-based agent adds an explicit goal and considers the future: it searches or plans over sequences of actions to find one that reaches the goal. This makes it flexible — change the goal and behavior changes — at the cost of more computation than reflex rules.
A utility-based agent goes beyond a binary goal by assigning a utility — a desirability score — to states, so it can choose among many goal-reaching paths and trade off conflicting objectives like speed versus safety. It picks the action with the highest expected utility.
A learning agent improves with experience. A learning element updates its behavior from feedback, a performance element chooses actions, a critic evaluates results against the performance measure, and a problem generator suggests exploratory actions. Any of the four types above can be made into a learning agent.
Ground the five types in familiar systems so the distinctions stick.
Simple reflex: a thermostat or a spam filter with fixed if-then rules. Model-based reflex: a robot vacuum that maintains a map of the room to cover it despite not seeing all of it at once. Goal-based: a GPS route planner that searches paths to reach a destination.
Utility-based: a self-driving car balancing safety, speed, comfort, and legality via a scoring function; or a recommender weighing relevance against diversity. Learning: a game-playing system like AlphaGo or a fraud detector that improves as it sees more outcomes.
Real systems mix types. A self-driving car is model-based (tracks the world), goal-based (has a destination), utility-based (trades off objectives), and learning (improves from data) at once. The five types are lenses, not exclusive boxes.
Connect the textbook taxonomy to today's LLM-powered agents.
A modern LLM-based agent fits the classic definition exactly: it perceives (reads a prompt, tool outputs, and retrieved data), decides (the model reasons about the next action), and acts (calls tools) in a loop toward a goal. It behaves like a goal-based, often utility-influenced, learning-capable agent.
What changed is the decision engine. Classic agents used hand-written rules, search, or explicit utility functions; an LLM agent uses a language model to choose actions from natural-language goals, which makes it far more general but also less predictable.
A plain classifier or a one-shot chatbot is not an agent — it perceives and outputs once with no goal-directed action loop. The agent label belongs to systems that act on their environment over multiple steps to pursue a goal.
An intelligent agent perceives its environment and acts on it to achieve goals, specified with PEAS and judged by a performance measure. Classical AI names five types by sophistication — simple reflex, model-based reflex, goal-based, utility-based, and learning — each adding capability over the last, and real systems combine them. Modern LLM agents fit the same perceive-decide-act definition, swapping hand-written rules for a language model as the decision engine.
Take a system you use — a robot vacuum, a recommender, or a self-driving feature. Write its PEAS, classify it among the five types (it may span several), and justify why it needs a world model, goals, or utility rather than simple reflex rules.
What is an intelligent agent in AI?
An intelligent agent maps percepts to actions to pursue a goal; it is rational when it chooses the action expected to best meet its performance measure.
What are the five types of intelligent agents?
Classical AI orders agents by sophistication: reflex rules, an internal world model, explicit goals, utility optimization, and learning from experience.
What is the difference between a goal-based and a utility-based agent?
Utility adds a desirability measure on top of a binary goal, letting the agent pick the best among many goal-reaching options and balance conflicting aims.
Are LLM-based agents the same as classic AI agents?
LLM agents are the classic goal-based, learning-capable agent with a new, more general but less predictable decision engine: the language model.