Two lists define what an intelligent agent is in AI. In a 1995 paper in The Knowledge Engineering Review, Michael Wooldridge and Nicholas Jennings named four properties: autonomy, social ability, reactivity, and pro-activeness. University courses based on Russell and Norvig's textbook add a standard for judging behavior, rationality, and a ladder of five designs from simple reflex agents to learning agents.
Key takeaways
- An agent perceives an environment through sensors and acts on it through actuators; course materials list humans, robots, software agents, and thermostats as examples.
- Wooldridge and Jennings call autonomy, social ability, reactivity, and pro-activeness the "weak notion" of agency.
- A rational agent picks the action expected to maximize its performance measure, given what it has perceived and what it already knows.
- The basic designs rise in generality from simple reflex to model-based, goal-based, and utility-based agents, and any of them can be turned into a learning agent.
- Having these characteristics does not make an agent human-like; the broad definition covers very simple programs.
What counts as an agent
Course lectures that follow the Russell and Norvig textbook start from a plain definition: an agent interacts with its environment, perceives it using sensors, and acts on it using actuators. Lecture slides from Michigan Technological University's CS 5811 course say agents include humans, robots, softbots, and thermostats, and describe the agent function as a mapping from the history of percepts to actions. A softbot is a software agent whose sensors and actuators are commands and replies in a software environment.
Under this definition, "intelligent" is a matter of how well the mapping from percepts to actions is chosen, not of whether the agent resembles a person. A thermostat is the simplest case. That breadth is deliberate, and it is the reason the characteristics below matter: they separate trivial agents from capable ones.
The four properties from the 1995 paper
The Wooldridge and Jennings paper appeared in volume 10, issue 2 of the journal, pages 115 to 152. It lists four properties that together make up what the authors call a weak notion of agency, a way of thinking of an agent as a kind of software process.
- Autonomy: the agent operates without direct intervention from humans or others and has some control over its own actions and internal state.
- Social ability: the agent interacts with other agents, and possibly people, through some kind of agent-communication language.
- Reactivity: the agent perceives its environment and responds in a timely way to changes in it.
- Pro-activeness: the agent does not only react; it shows goal-directed behavior by taking the initiative.
The paper also describes a stronger notion, used by some AI researchers, in which an agent is described with human concepts such as knowledge, belief, and intention. Attributes such as mobility, veracity, benevolence, and rationality are discussed separately as optional extras.
Reading the properties on everyday agents
The table is an illustrative reading of the four properties for three agents. The judgments are the author's, not a ranking from the paper, and a different designer could rate the middle row differently.
| Example agent | Autonomy | Reactivity | Pro-activeness | Social ability |
|---|---|---|---|---|
| Wall thermostat | Yes, runs unattended | Yes, responds to temperature | No, only responds | No |
| Robot vacuum with a schedule | Yes | Yes, reacts to obstacles | Yes, starts cleaning on its own initiative | Limited, talks to a phone app |
| Delivery-routing software that bargains with other planners | Yes | Yes | Yes | Yes, exchanges messages with other agents |
Only the last row has all four, which is why research on agents that negotiate or coordinate treats social ability as a separate property.
Rational, not all-knowing
The characteristic that connects agents to decision-making is rationality. A lecture in the University of Waterloo's CS 486/686 course gives the textbook definition: for each possible percept sequence, a rational agent should select an action that is expected to maximize its performance measure, given the evidence from the percept sequence and whatever prior knowledge the agent has.
Three points follow from the wording. The standard is expected performance, so a rational choice can still turn out badly when the world is random. It is relative to a performance measure someone chose, so a thermostat that minimizes energy and one that maximizes comfort are rational about different things. And it depends on what the agent has perceived, so an agent is not blamed for facts it could not sense.
Five designs, shown on one heating controller
The Michigan Tech slides list four basic agent types in order of increasing generality and complexity, simple reflex, reflex with state, goal-based, and utility-based, and note that all can be turned into learning agents. The example below is illustrative: one home heating controller, with invented numbers, redesigned at each step.
| Design | What the controller does | What it adds |
|---|---|---|
| Simple reflex | If the room reads below 20 degrees Celsius, turn the heat on | Condition-action rule on the current reading only |
| Reflex with state | Remembers recent readings and knows an open window makes the temperature fall quickly | An internal model of how the world changes |
| Goal-based | Starts heating early enough for the room to reach 21 degrees Celsius by 7 a.m. | A target state and a look-ahead to reach it |
| Utility-based | Weighs comfort against electricity cost when a cheaper, slightly cooler plan exists | A score for how good each outcome is |
| Learning | Adjusts its start time after seeing how long mornings actually took | Improvement from experience |
Each step fixes a weakness of the one above it. The first design fails when one reading is misleading, the second cannot plan ahead, the third cannot choose between two ways of reaching the goal, and the fourth starts with a fixed idea of the house that may be wrong.
Common misreadings
- A system with these characteristics is not necessarily general-purpose; a thermostat has most of them.
- Autonomy describes control over its own actions, not freedom from rules or from human oversight.
- Rational does not mean correct in hindsight, and it does not mean human-like reasoning.
FAQ
What are the main characteristics of an intelligent agent?
The most cited list is autonomy, social ability, reactivity, and pro-activeness from Wooldridge and Jennings in 1995. Textbook treatments add rationality and the choice of design, from simple reflex to learning.
Is a thermostat really an agent?
Under the textbook definition, yes. It senses temperature and acts on a heater. It is a simple reflex agent, lacking pro-activeness and social ability.
What separates a goal-based agent from a utility-based one?
A goal-based agent aims at a target state and treats reaching it as success. A utility-based agent scores outcomes, so it can choose among several ways of reaching the goal or trade one aim against another.
Does an agent have to learn?
No. Learning is an optional addition that the course slides say can be applied to any of the basic types.
Related reading
- How AI differs from machine learning
- Predictive analytics in practice
- How AI handles complex problems
- AI that builds AI
More in AI Basics.
Last verified: October 2026
Sources
- Wooldridge M, Jennings NR, Intelligent agents: theory and practice, The Knowledge Engineering Review 10(2), 115-152, 1995
- Onder N, CS 5811 Advanced Artificial Intelligence, lecture slides on chapters 1 and 2, Michigan Technological University, 2020
- Gao A, CS 486/686 Introduction to Artificial Intelligence, Lecture 2: Rational agents, University of Waterloo, 2019
- Russell S, Norvig P, Artificial Intelligence: A Modern Approach, 4th US edition, companion site, University of California, Berkeley