Intelligent Agents

An intelligent agent is any entity that perceives its environment through sensors and acts upon it through actuators in order to achieve goals. This definition, popularized by Stuart Russell and Peter Norvig in «Artificial Intelligence: A Modern Approach», runs through much of the field: rather than asking what thinking is, AI asks what a situated entity should do to behave well.

The central idea is the rational agent: the one that, given the information available to it, chooses the action expected to maximize its performance measure. Rationality is not omniscience or perfection; it is doing the best one can with what is perceived. A thermostat, a route planner and a conversational assistant are all agents insofar as they close that perceive-decide-act loop.

Types of agent by complexity

Russell and Norvig arrange agents on a scale of increasing sophistication. Simple reflex agents apply condition-action rules to the current percept, with no memory. Model-based reflex agents keep an internal state that stands in for the part of the world they cannot observe directly. Goal-based agents choose actions that lead to desired states, which requires reasoning about the future. Utility-based agents weigh how good each outcome is through a utility function, which matters when goals conflict or outcomes are uncertain. Finally, learning agents improve their behavior with experience, tuning any of the previous components from feedback. This is not a hierarchy of quality but of resources: each level handles environments the previous one cannot.

The properties of the environment

Designing an agent depends as much on the environment as on the agent, and Russell and Norvig classify that environment along several dimensions. It is fully observable if the sensors capture everything relevant, or partially observable if the hidden parts must be inferred. It is deterministic when the action fixes the next state, or stochastic when chance intervenes. It is static if it does not change while the agent deliberates, or dynamic if the world does not wait. And it is discrete or continuous depending on the nature of its states, percepts and actions. Chess is observable, deterministic and discrete; driving a car is partial, stochastic, dynamic and continuous. The more demanding the environment, the more sophisticated the agent that lives in it must be.

From classic agents to today's LLM agents

In 2026 the word «agent» also names something new: agents built on large language models. Such a system takes a goal in natural language, plans a sequence of steps and acts using tools -a browser, a code interpreter, APIs-, observes the result of each step and corrects in a loop until it reaches the goal or exhausts its budget. The underlying definition has not changed: it is still perceive, decide and act. What LLMs contribute is the policy, the criterion that picks the next move, learned from data rather than hand-coded; they fit the classic taxonomy as goal-based agents that learn. What remains unresolved is their reliability and their safe degree of autonomy: they hallucinate, they compound errors, and they still require human oversight on critical tasks. For their practical application, this glossary points to the «AI agents» entry. Documentation: official Artificial Intelligence: A Modern Approach material; original ReAct paper.

Autonomy, privacy and security

When an agent can call tools, a wrong output is no longer just text: it can become an action affecting email, files, money or accounts. OWASP calls this risk excessive agency when functionality, permissions or autonomy exceed what the task requires, with potential effects on confidentiality, integrity and availability. Privacy fails when context or a tool exposes data the task did not need; security fails when manipulated input succeeds in ordering an improper action. The control is not to ask the model to be careful: minimize tools and permissions, authorize each operation in the downstream system and require human approval for high-impact decisions.