Intelligent Agents and Environments #

An intelligent agent connects perception with action. It observes an environment through sensors, uses the information it receives to decide what to do, and affects the environment through actuators.

This page develops the core vocabulary for reasoning about intelligent agents: percepts, actions, rationality, performance measures, PEAS and the different properties an environment can have.

Learning Objectives #

By the end of this page, you should be able to:

• explain the relationship between an agent and its environment
• distinguish sensors, percepts, actions and actuators
• explain percept sequences and the agent function
• describe what makes an agent rational
• distinguish rationality from omniscience
• specify a task environment using PEAS
• classify environments using standard AI properties
• explain why environment properties affect agent design

Big Picture #

flowchart TD
    E[Environment] -->|Percepts| S[Sensors]
    S --> A[Agent]
    A --> C[Actuators]
    C -->|Actions| E

    style E fill:#E1F5FE
    style S fill:#FFF9C4
    style A fill:#C8E6C9
    style C fill:#EDE7F6

Think of an agent as a continuous loop: observe → decide → act → observe again. An action may change the environment, so the next observation may be different.

1. What Is an Agent? ☆ #

The definition is deliberately broad.

An agent does not exist in isolation. Its sensors and actuators only make sense relative to the environment in which it operates.

Agent and environment #

The environment is the part of the world relevant to the problem the agent is trying to solve.

For a navigation agent, the environment might contain roads, possible locations, obstacles and a destination. For a medical diagnosis system, it might include the patient, available clinical information and the healthcare setting.

The exact environment therefore depends on the problem being modelled.

2. Sensors, Percepts, Actions and Actuators ☆ #

These terms describe different parts of the agent-environment interaction.

For a self-driving vehicle:

• a camera is a sensor
• the observed lane, pedestrian or traffic signal contributes to the percept
• deciding to slow down is an action
• the braking system is an actuator that carries out the action

Do not confuse a sensor with a percept. A sensor is the mechanism that observes; a percept is the information obtained through observation.

3. Percepts and Percept Sequences ☆ #

A percept represents what an agent observes at a particular moment.

A percept sequence is the complete history of percepts the agent has received so far.

This distinction matters because some decisions depend only on the current observation, while others require information from earlier observations.

For example, an agent navigating through a changing environment may need to remember where it has already been rather than deciding solely from the current view.

Vacuum-cleaner example #

Consider a simple vacuum-cleaner world with two locations, A and B.

A percept might contain:

[A, Dirty]

This tells the agent:

• its current location is A
• location A is dirty

Possible actions include:

Left, Right, Suck, NoOp

The agent uses the percept to select an appropriate action. If it perceives [A, Dirty], for example, Suck may be an appropriate choice.

4. Agent Function ☆ #

The agent function describes the mapping from percept history to an action.

where:

• \( P^{*} \) represents possible percept sequences
• \( A \) represents the set of possible actions

In simple words:

percept history → agent function → action

The agent program is the implementation that runs on a physical or software architecture to realise this behaviour.

The conceptual function describes what action should be selected; the program and architecture determine how that behaviour is implemented.

5. Rational Agents ☆ #

A rational agent should try to do the right thing based on:

• what it has perceived
• what it already knows
• the actions available to it
• how success is measured

The word expected is important. An agent may have to act under uncertainty, where no action guarantees the best outcome.

Performance measure #

A performance measure is an objective criterion used to judge how successfully an agent behaves.

For a vacuum-cleaner agent, performance might consider:

• amount of dirt cleaned
• time taken
• electricity consumed
• noise generated

There may therefore be more than one measure of success. A system that cleans everything but uses excessive energy may not be considered optimal when energy efficiency is also important.

Rationality is not omniscience #

A rational agent is not required to know the future.

Omniscience would mean knowing the actual outcome of every possible action in advance. Real agents do not have this ability.

Rationality instead asks whether the agent made the best decision it could using the evidence available at the time.

A rational decision can still lead to a poor outcome. If the future is uncertain, the quality of the decision should be judged using the information available when the decision was made, not information revealed afterwards.

Exploration and information gathering #

A rational agent may deliberately take an action that improves its future knowledge. It can explore the environment, gather information and use new percepts to make better decisions later.

Autonomy #

An agent is more autonomous when its behaviour is determined by its own experience rather than being completely fixed by built-in knowledge.

Learning and adaptation therefore support autonomy.

6. Bounded Rationality #

Perfect rationality is usually unrealistic because an agent has limited:

• computational power
• memory
• time
• knowledge of the environment

Bounded rationality recognises these limits. The aim is to choose the best practical action given the agent’s knowledge and available computational resources.

flowchart TD
    A[Performance objective] --> B[Available knowledge]
    B --> C[Possible actions]
    C --> D[Resource limits]
    D --> E[Best practical action]

    style A fill:#E1F5FE
    style B fill:#FFF9C4
    style C fill:#EDE7F6
    style D fill:#FFF9C4
    style E fill:#C8E6C9

7. PEAS - Specifying a Task Environment ☆ #

Before designing an intelligent agent, we need to describe the task it is expected to perform and the world in which it operates.

PEAS provides a systematic way to do this:

Remember PEAS as: How is success measured? Where does the agent operate? How can it act? How can it sense?

Worked example: automated taxi #

The performance measure is deliberately broader than simply “reach the destination”. A taxi that arrives quickly but drives dangerously is not behaving successfully.

Medical diagnosis system #

Part-picking robot #

The PEAS description changes with the task. It should therefore be defined from the problem statement, not copied from another agent.

8. Properties of Environments ☆ #

The nature of the environment strongly influences how an intelligent agent should be designed.

Important dimensions include:

• fully observable vs partially observable
• deterministic vs stochastic
• episodic vs sequential
• static vs semi-dynamic vs dynamic
• discrete vs continuous
• single-agent vs multi-agent

Fully observable vs partially observable #

A fully observable environment gives the agent access through its sensors to all information relevant to choosing an action at a particular time.

A partially observable environment hides some relevant information or provides incomplete/noisy observations.

Examples:

• chess is typically treated as fully observable because the complete board state is visible
• poker is partially observable because an agent cannot see every player’s cards

Deterministic vs stochastic #

In a deterministic environment, the next state is completely determined by the current state and the action taken.

In a stochastic environment, there is uncertainty about the next state.

Examples of stochastic settings include autonomous driving, a robot playing football, and agricultural tasks affected by changing weather.

If uncertainty arises specifically from the actions of other agents, the environment can also be described as strategic.

Episodic vs sequential #

In an episodic environment, each decision can be treated largely independently. The current action does not depend on a long history of earlier decisions.

Example: classifying whether an individual image contains a cat.

In a sequential environment, current actions affect future states and decisions. Long-term consequences therefore matter.

Examples include navigation, games and investment portfolio management.

Static, dynamic and semi-dynamic #

A static environment does not change while the agent is deciding what to do.

Examples include many puzzles and a Rubik’s cube while no external actor changes it.

A dynamic environment can change while the agent is deliberating.

Examples include road traffic and changing weather.

A semi-dynamic environment does not change while the agent thinks, but the agent’s performance changes with time. Chess played with a clock is a common example.

Discrete vs continuous #

A discrete environment has distinct states, percepts or actions that can be represented as separate choices. Chess is a standard example.

A continuous environment contains quantities that vary over a range, such as position, speed, steering angle or time. Taxi driving therefore contains many continuous elements.

Single-agent vs multi-agent #

A single-agent environment contains one decision-making agent of interest.

A multi-agent environment contains other agents whose actions can influence what happens.

Multi-agent relationships may be:

• cooperative
• competitive
• self-interested

Game playing is an obvious example, but many real-world environments are also multi-agent because people, vehicles or software systems interact with one another.

9. Environment Classification Examples #

The real world is commonly partially observable, stochastic, sequential, dynamic, continuous and multi-agent. This combination makes real-world agent design substantially harder than simple toy environments.

10. Why Environment Properties Matter #

Environment classification is not just terminology. It guides the design of the agent.

For example:

• partial observability may require memory or belief about hidden state
• stochastic environments require reasoning under uncertainty
• sequential tasks require consideration of future consequences
• dynamic environments reward timely decisions
• continuous environments need methods that can handle continuous quantities
• multi-agent environments require reasoning about the actions of others

flowchart TD
    A[Environment properties] --> B[What can be observed?]
    A --> C[How does the world change?]
    A --> D[Do actions affect the future?]
    B --> E[Agent design]
    C --> E
    D --> E

    style A fill:#E1F5FE
    style B fill:#FFF9C4
    style C fill:#EDE7F6
    style D fill:#FFF9C4
    style E fill:#C8E6C9

11. AI Agent and Agentic AI - A Useful Distinction #

The broad AI definition of an agent is older and more general than the recent use of the phrase agentic AI.

In the classical sense used here, an agent is defined by perception and action in an environment. It could be a robot, a software program or another decision-making system.

The term agentic AI is often used for systems with a higher degree of autonomy, where one or more AI components can reason about goals, break work into tasks, use tools, coordinate actions and adapt their behaviour.

The important point for this topic is that the theory of intelligent agents does not depend on large language models or modern tool-calling systems.

Common Mistakes #

• A sensor is not a percept: the sensor obtains information; the percept is the information received.
• An actuator is not the same as an action: the action is the decision; the actuator carries it out.
• Rational does not mean all-knowing. Rationality is judged using the information available to the agent.
• A task may have several performance criteria; success is rarely captured by a single number in realistic systems.
• Environment properties are not universal labels for an agent. They depend on the particular task and modelling assumptions.

Practice Questions #

1. Define an agent, sensor, percept, action and actuator using one continuous example.
2. What is the difference between a percept and a percept sequence?
3. Explain what the agent function \( f : P^{*} \rightarrow A \) represents.
4. Why is rationality different from omniscience?
5. Give four possible performance measures for a vacuum-cleaner agent.
6. Construct a PEAS description for an autonomous delivery robot.
7. Distinguish deterministic and stochastic environments with examples.
8. Why is image classification usually treated as episodic while navigation is sequential?
9. What is the difference between static, dynamic and semi-dynamic environments?
10. Classify an autonomous taxi environment across the major environment properties and justify each choice.

Key Takeaways #

• An agent perceives an environment through sensors and acts through actuators.
• A percept is one observation; a percept sequence is the history of observations.
• The agent function maps percept sequences to actions.
• A rational agent selects actions expected to maximise its performance measure using the information available.
• PEAS = Performance measure, Environment, Actuators, Sensors.
• Environment properties determine what information and decision-making capabilities an agent needs.

Checklist #

• I can explain the agent-environment interaction.
• I can distinguish sensors, percepts, actions and actuators.
• I can explain the agent function and percept sequence.
• I can distinguish rationality from omniscience.
• I can construct a PEAS description for a new problem.
• I can classify an environment by observability, determinism, dependency, dynamics, continuity and number of agents.
• I can explain why environment properties influence agent design.

Home | Artificial and Computational Intelligence