An intelligent agent is anything that perceives its environment through sensors and acts upon that environment through actuators in order to achieve goals. In artificial intelligence, the concept offers a common way to describe very different systems, from a thermostat to a robot vacuum, a chess program or a self-driving car. Many textbooks, including the widely used Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig, organize the whole field around the idea of building agents that act rationally. A rational agent is one that chooses the action expected to give the best result according to a measure of success, given what it knows.
History of the intelligent agent concept
The word agent has long been used in computing for programs that act on a user’s behalf. In artificial intelligence, the agent view gained ground in the 1980s and 1990s as researchers moved from isolated problem-solving programs to systems that operate in changing, uncertain surroundings. Russell and Norvig made it the organizing idea of their textbook, first published in 1995, and the framing spread through teaching and research. Related fields such as economics and control theory had already studied decision makers choosing actions to maximize some payoff. After 2022, the term returned to wide public use when developers began to build assistants on top of large language models that can plan and take steps using software tools, often called AI agents.
How an intelligent agent works
The basic loop is simple. The agent receives a percept, which is whatever its sensors report at a moment in time. It combines the percept with what it has stored from the past, decides on an action and carries it out through actuators. The action changes the environment, producing new percepts, and the cycle repeats. A robot vacuum senses walls and dirt with bumpers and cameras and acts with wheels and a brush. A software agent might sense email messages and act by filing them. The part that maps percepts to actions is often called the agent function, and the program that implements it is the agent program.
The PEAS framework
To specify an agent clearly, designers often describe four elements, summarized as PEAS.
- Performance measure: how success is judged, such as safety, speed and passenger comfort for a taxi.
- Environment: the setting in which the agent operates, such as roads, traffic and weather.
- Actuators: the means of acting, such as steering, braking and a horn.
- Sensors: the means of perceiving, such as cameras, radar and GPS.
Writing down these four items before designing a system forces clarity about what the agent is for and what it can do. The performance measure is especially important, because an agent will pursue whatever it is measured on, even when that is not exactly what its designers meant.
Types of intelligent agents
The standard classification orders agents by increasing capability.
- Simple reflex agents choose actions from the current percept alone, using condition-action rules. A thermostat that turns on heat below a set temperature is an example.
- Model-based reflex agents keep an internal model of the parts of the world they cannot see directly, so they can handle partial information.
- Goal-based agents consider the future consequences of actions and choose those that lead toward a stated goal, which often requires search and planning.
- Utility-based agents go further by assigning a numerical value to outcomes, letting them weigh trade-offs such as speed against risk.
- Learning agents improve their performance with experience, using a learning element that changes how they act.
Many real systems combine these designs. A modern robot may use reflexes for balance, a model of the room, goals for tasks and learning to improve over time.
Properties of environments
The difficulty of building an agent depends heavily on its environment. In a fully observable environment, the agent can see everything relevant, while in a partially observable one it must infer hidden facts. Environments can be deterministic, where actions have predictable results, or stochastic, where chance plays a role. They can be episodic, with independent decisions, or sequential, where each choice affects later ones. They may be static or dynamic, discrete or continuous, and involve a single agent or many. Chess is fully observable and deterministic, while driving in city traffic is partially observable, stochastic, sequential, dynamic and multi-agent. Harder environments demand more capable designs.
Intelligent agents and AI agents today
In current usage, AI agent often refers to a system built around a large language model that can break a goal into steps, call tools such as search engines or calculators and keep a record of what it has done. This fits the classical definition. The language model supplies reasoning and language skills, the tools act as actuators and the text or data it receives acts as percepts. The same concerns also apply: the agent must be given a clear goal, kept within safe limits and monitored, because an agent that acts in the world can cause real consequences if it misjudges a situation.
Applications of intelligent agents
Intelligent agents are used in robotics, where machines navigate and manipulate objects, and in games, where computer opponents act within rules. They run in finance as trading programs, in customer service as virtual assistants, in logistics as route planners and in smart homes as energy controllers. Multi-agent systems, in which many agents interact, model markets, traffic and teams of robots. In each case the agent framework helps engineers state the goal, the available actions and the information the system has.
Challenges in designing intelligent agents
Building a capable agent raises several recurring difficulties. The agent must act with incomplete information, because sensors are noisy and some facts are hidden. It must balance short-term gains against long-term goals and decide how much to explore new actions and how much to repeat those known to work. It must also cope with other agents whose behavior may be cooperative, competitive or unpredictable. Computing the best action can be expensive, so real agents usually settle for good enough choices within a time limit, a idea known as bounded rationality. Finally, specifying the goal correctly is hard. A badly chosen performance measure can lead an agent to exploit loopholes, such as a cleaning robot that hides dirt instead of removing it.
Safety and responsibility
As agents gain more autonomy, questions of safety and accountability grow. Designers set limits on what actions an agent may take, require human approval for important steps and keep logs so that behavior can be reviewed. Testing in simulation before release helps catch dangerous errors. For agents that act through the internet, security matters too, since an attacker may try to trick the agent with hidden instructions. Standards bodies and regulators increasingly expect that organizations deploying agents document their purpose, monitor performance and provide a way for people to correct or stop them.
FAQs
Is a thermostat an intelligent agent? Under the broad textbook definition, yes, though it is a very simple reflex agent. Many people reserve the word intelligent for systems that plan or learn.
What is a rational agent? It is an agent that picks the action expected to maximize its performance measure, given its percepts and built-in knowledge.
How do intelligent agents differ from ordinary programs? They are defined by sensing an environment and acting on it toward goals, often with some autonomy and the ability to adapt.
More in Artificial Intelligence
Key facts
- In the standard textbook definition, an agent perceives its environment through sensors and acts upon it through actuators.1
- A rational agent selects the action expected to maximize its performance measure, given what it has perceived.1
- The textbook classifies agents as simple reflex, model-based reflex, goal-based, utility-based and learning agents.2
Related concepts
In the news
Quick checkWhat are the five classic types of intelligent agent?Show answer
Simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents.
Sources
- Russell and Norvig, Artificial Intelligence: A Modern Approach, https://aima.cs.berkeley.edu/
- Wikipedia, Intelligent agent, https://en.wikipedia.org/wiki/Intelligent_agent
- IBM, What are AI agents?, https://www.ibm.com/think/topics/ai-agents
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
