AI agent

What an agent is

An intelligent agent is an entity that perceives its environment, acts autonomously to achieve goals, and improves its performance by acquiring knowledge^[1]. The definition is deliberately broad: Russell and Norvig, who gave it in 2003, placed a thermostat and a human being in the same category^[1]. There is no single standard definition of an agent, and different fields use the word in different ways^[1].

Inside AI, the usual refinement is the rational agent, which chooses the action expected to maximise its expected utility, as framed by Russell and Norvig in 2021^[1].

How agents work

Agents perceive through sensors and act through actuators. What an agent does is described by an agent function, which maps a sequence of percepts to an action^[1]. What counts as a sensor or an actuator depends on the agent: a physical device in one case, a software interface in another.

A concrete case is the simple reflex agent, which behaves like a lookup table. For every percept it may encounter, a matching action is prepared in advance, and it simply performs the action that goes with what it currently perceives. The difficulty is that environments are often only partially observable: the same percept can recur in a situation where the rule fires again, so such agents can fall into infinite loops^[1].

Kinds of agents

Russell and Norvig classify architectures into five kinds: simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents^[1]. Weiss in 2013 offered a different split: logic-based or deliberative agents, reactive agents, belief-desire-intention agents, and layered or hybrid agents^[1]. The belief-desire-intention architecture comes from Rao and Georgeff in 1995^[1].

Workflows and agents

Recent usage has added the label agentic AI, which refers to systems that decide autonomously, act over longer horizons and include memory and orchestration software^[1]. Examples cited include AWS Amazon Connect Health and Salesforce Agentforce in 2026^[1].

Anthropic groups all of these variations as agentic systems but draws an architectural distinction between workflows and agents^[2]. Workflows are systems where LLMs and tools are orchestrated through predefined code paths^[2]. Agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks^[2].

The practical consequence of that distinction is a recommendation to find the simplest solution possible and only increase complexity when needed, which may mean not building an agentic system at all^[2].

Limits

The word covers a very wide range of things. Because the 2003 definition stretches from a thermostat to a person, and because there is no single standard definition, statements about what an agent is or does are hard to compare across contexts^[1]. The classic architectures also carry known failure modes, such as reflex agents looping in partially observable environments^[1]. And the recent, looser usage bundles together systems with very different amounts of autonomy, which is why the workflow-versus-agent split is useful: it asks where control actually sits, in predefined code or in the model itself^[2].