Intelligent Agents

An intelligent system does not merely answer questions.

It exists in an environment.

It perceives.

It chooses.

It acts.

This is the central idea behind the intelligent agent framework.

Agent

An agent is a system that:

• receives percepts,
• selects actions.

Its actions affect the environment.

This creates a feedback loop.

Agent–Environment Loop

A simple cycle is:

[ Environment \rightarrow Perception \rightarrow Agent \rightarrow Action \rightarrow Environment ]

The process repeats.

Intelligence unfolds through interaction.

Sensors

An agent perceives through sensors.

Examples include:

• camera,
• microphone,
• keyboard input,
• software API.

A sensor is any channel carrying information into the agent.

Actuators

An agent acts through actuators.

Examples:

• wheels,
• robotic arms,
• text output,
• database commands.

An actuator changes the environment.

Percept

A percept is the information received at one moment.

A percept sequence is the history of what the agent has observed.

Good action may depend on the whole history.

Agent Function

An abstract agent function maps percept history to action:

[ f:P^*\rightarrow A ]

where:

• (P^*) = possible percept sequences,
• (A) = actions.

This is a formal definition of behavior.

Agent Program

The abstract function is implemented by an agent program running on some architecture.

Again:

function

and:

physical realization

are distinct.

Rational Agent

A rational agent chooses actions expected to maximize its performance measure given:

• percept history,
• knowledge,
• available actions.

Rationality is relative to information.

Performance Measure

The performance measure defines success.

For a delivery robot, it might include:

• speed,
• safety,
• energy use.

Poorly chosen metrics create poor behavior.

PEAS

A classic AI framework describes task environments using PEAS:

• Performance measure,
• Environment,
• Actuators,
• Sensors.

It forces explicit specification.

Example: Autonomous Taxi

Performance:

• safety,
• legality,
• travel time.

Environment:

• roads,
• traffic,
• pedestrians.

Actuators:

• steering,
• brakes,
• accelerator.

Sensors:

• cameras,
• radar,
• maps.

Environment Properties

AI classifies environments along several dimensions.

These classifications determine what agent architecture is appropriate.

Fully Observable vs Partially Observable

In a fully observable environment, the agent has access to the relevant state.

Chess is close to this.

Driving is partially observable.

Some information is hidden.

Deterministic vs Stochastic

In deterministic environments, an action has predictable effects.

In stochastic environments, outcomes are uncertain.

Real-world domains are often stochastic.

Episodic vs Sequential

In an episodic task, each decision is largely independent.

In sequential tasks, current actions affect future situations.

Chess and driving are sequential.

Static vs Dynamic

A static environment waits while the agent thinks.

A dynamic environment changes during deliberation.

Real-time systems must account for time.

Discrete vs Continuous

Chess has discrete moves.

Driving involves continuous:

• position,
• velocity,
• time.

Different mathematics applies.

Single-Agent vs Multi-Agent

Some environments contain one decision maker.

Others contain:

• competitors,
• collaborators.

Multi-agent settings require strategic reasoning.

Simple Reflex Agent

A simple reflex agent uses condition–action rules.

Example:

IF obstacle ahead THEN turn.

It reacts only to current percept.

Limitation of Reflexes

If the environment is partially observable, current percept may be insufficient.

The agent needs memory.

Model-Based Agent

A model-based agent maintains internal state.

It estimates aspects of the world not directly visible.

This connects to mental models.

Goal-Based Agent

A goal-based agent evaluates actions relative to desired states.

It can plan.

Goals introduce future structure.

Utility-Based Agent

A utility-based agent ranks outcomes.

This helps when several goals conflict.

Utility supports tradeoffs.

Learning Agent

A learning agent improves through experience.

A classical architecture includes:

• performance element,
• learning element,
• critic,
• problem generator.

Performance Element

The performance element chooses actions.

It embodies current competence.

Learning Element

The learning element modifies the system to improve future behavior.

It changes policy or knowledge.

Critic

The critic evaluates outcomes.

It provides feedback.

Without evaluation, learning lacks direction.

Problem Generator

The problem generator encourages exploration.

It suggests actions that may yield useful new experience.

Learning requires experimentation.

Autonomy

An autonomous agent relies increasingly on its own experience rather than only built-in knowledge.

Autonomy is gradual.

No real system is completely independent.

Model of the World

Agents often need an internal model representing:

• objects,
• dynamics,
• other agents.

The model supports prediction.

Model-Based Planning

If the agent knows:

[ P(s’|s,a) ]

it can simulate consequences of actions.

Planning becomes model-based control.

Policy

A policy specifies action selection:

[ \pi(a|s) ]

It may be:

• deterministic,
• stochastic.

The policy is the agent’s behavioral strategy.

Reward

In reinforcement learning, reward provides scalar feedback.

The agent learns which actions lead to high long-term return.

Reward does not need to specify exactly how to behave.

Partial Observability

When state is hidden, the agent may maintain a belief state.

A belief state is a probability distribution over possible world states.

Decision making becomes inference under uncertainty.

Belief Update

After observing evidence, the agent updates its belief.

This is conceptually Bayesian.

Perception becomes state estimation.

Multi-Agent Systems

When several agents interact, behavior depends on others.

They may:

• cooperate,
• compete,
• negotiate.

Game theory becomes relevant.

Cooperative Agents

Agents can share:

• information,
• goals,
• tasks.

Coordination can create collective intelligence.

Competitive Agents

An adversary actively tries to defeat the agent.

Game playing is a clear example.

The environment becomes strategic.

Communication

Multi-agent systems often need protocols.

Communication can improve coordination.

But messages can be:

• costly,
• deceptive,
• incomplete.

Social Agents

Human-facing agents must interpret:

• norms,
• intentions,
• expectations.

Social intelligence extends beyond task optimization.

Tool-Using Agents

Software agents can interact with external tools such as:

• search,
• calculators,
• code execution.

The agent becomes an orchestrator.

Memory

A capable agent often needs several kinds of memory:

• working state,
• episodic history,
• semantic knowledge.

Memory supports long-horizon behavior.

Planning Horizon

Short-horizon agents optimize immediate consequences.

Long-horizon agents consider distant outcomes.

Long horizons increase computational difficulty.

Receding-Horizon Control

A practical strategy is:

plan several steps ahead, act once, observe, replan.

This balances foresight with uncertainty.

Hierarchical Agents

Complex tasks can be decomposed.

A high-level agent chooses subgoals.

Lower-level controllers execute them.

Hierarchy reduces complexity.

Options

In reinforcement learning, options are temporally extended actions.

Instead of choosing every primitive action, the agent can choose:

“navigate to door.”

This supports abstraction.

Agentic Failure

An agent can fail because of:

• bad perception,
• bad model,
• bad planning,
• bad objective.

Intelligence is architectural.

One strong component cannot compensate for every weak one.

Specification Risk

If the performance measure is incomplete, the agent may optimize the wrong behavior.

This is especially serious for autonomous systems.

Instrumental Convergence

Some theorists note that many goals may create similar useful subgoals such as:

• acquiring resources,
• preserving capability.

This is called instrumental convergence.

Its significance depends on agent design and assumptions.

Agent vs Tool

A calculator is primarily a tool.

An agent chooses actions autonomously based on state and goals.

The boundary is gradual.

Modern AI increasingly blurs it.

Human as Agent

Humans fit the agent abstraction:

• sense,
• model,
• act.

But the abstraction omits:

• emotion,
• embodiment,
• social identity.

Models simplify.

Agents and Minds

An agent need not have a mind in the rich philosophical sense.

A thermostat can be described as a minimal agent.

Agency comes in degrees.

Agency and Consciousness

An artificial agent can pursue goals without evidence of subjective experience.

Agency is not consciousness.

The concepts should remain separate.

The Philosophical Lesson

The intelligent-agent framework turns intelligence into an interaction problem.

A system is intelligent insofar as it can:

• perceive,
• model,
• choose,
• act,
• learn

in an environment.

Intelligence is not stored inside the agent alone.

It emerges in the loop between agent and world.

The Next Question

How does an agent improve from experience?

Instead of encoding every rule manually, can it learn patterns directly from data?

That is the central question of:

Machine Learning.