What are intelligent agents, and how are they used in AI?

In this article

1. Before you start
2. Why interviewers ask this
3. Intelligent agents from first principles
4. Worked example
5. Trade-offs and edge cases
6. How to answer it in an interview
7. Common mistakes
8. Follow-up questions
9. Sources

Before you start

Most artificial intelligence (AI) agents people meet today are built around a prompted model, so prompt engineering is the one topic worth knowing first.

Why interviewers ask this

"Agent" now means almost anything, so a candidate who cannot say what it means shows they have only used the word. The question separates people who can draw a line between a program that runs and one that decides. It also leads quickly to design questions: when to give a model control, and when to keep a fixed path.

A weak answer says "an agent is an AI that does tasks for you." A strong answer gives the classic definition, names its four properties, and says that a program which only responds to input falls short of it. A Staff-level answer adds the engineering consequence: more freedom costs latency and money, so you add it only when the task needs it. The sources in this page back each of those points, and the agentic AI interview questions page collects the neighbouring prompts.

Intelligent agents from first principles

Intelligent agent

An intelligent agent is a computer system, situated in an environment, that is capable of flexible, autonomous action to achieve its design objectives, shown by autonomy, reactivity, proactiveness and social ability rather than by reacting alone.

Agentic system

An agentic system is an application class that combines an LLM with tools to finish a task; Anthropic uses the term for the whole family and splits it into workflows and agents.

In plain words

Picture a house-sitter you hired for a week. You do not tell them each action. They notice a dripping tap and fix it. They water the plants before the dry weekend, not after. They phone the plumber when the job is too big. Those are the four properties: working alone, noticing, planning ahead, and talking to others.

The analogy breaks in one place. A human sitter understands your intent. A software agent has only the goals and tools its designers gave it.

In one precise sentence, the Institute of Electrical and Electronics Engineers (IEEE) Technology Navigator describes an intelligent agent as a computational system situated in an environment that is capable of flexible, autonomous action to achieve design objectives. Jennings and Wooldridge (1998) word it nearly the same way. The environment is whatever the agent senses and changes, and it may be the physical world, a user, a collection of agents or the Internet.

Step 1: Autonomy: the agent acts without step-by-step control

Autonomy is the ability to act on its own, without a person approving every move. The agent is given an objective and chooses actions to pursue it. A thermostat that switches a heater on and off is autonomous in this narrow sense: nobody presses the button each time.

Input: an objective and the current situation. Output: an action. Autonomy alone is a weak test, because a timer is autonomous too. That is why the classic definition asks for flexible autonomous action, and the next three steps are what flexible means. Four properties (autonomy, reactivity, proactiveness and social ability) are credited to Wooldridge and Jennings in 1995 by the IEEE entry. You now know the base property: the agent acts on its environment without being steered. Step 2 adds what it must notice first.

Step 2: Reactivity: the agent senses its environment and responds

Reactivity is the ability to perceive the environment and respond to changes in it in good time. Jennings and Wooldridge (1998) call this property responsive: the agent perceives its environment, which may be the physical world, a user, a collection of agents or the Internet, and responds. A recent critical analysis puts it the same way: reactivity implies that the agent perceives its environment and responds to changes in a timely manner (arXiv 2506.01463, v1).

Input: what the agent senses, such as a temperature reading, a user message or a tool result. Output: a changed action. The loop then repeats, because the action changes the environment and the agent senses the new state. A simple reflex agent lives entirely in this step. You now know how an agent stays in touch with a changing world. Step 3 is the property most programs lack.

Step 3: Proactiveness: the agent works toward a goal

Proactiveness is the ability to take the initiative and act toward a goal, not only in response to the environment. Jennings and Wooldridge (1998) say that agents should not simply act in response to their environment. A program that only reacts waits until the situation is already bad. A proactive agent acts now because of what it expects later.

Input: a goal plus knowledge of what is coming, such as a calendar. Output: an action that helps a future state. The goal-based agent page covers how such an agent plans. The worked example below runs the same room with and without this property.

This is the error from the lead. Under the Wooldridge–Jennings definition, "it responds to input" makes a program an agent at most, not an intelligent agent, because the same source says flexible behaviour needs proactive and social behaviour as well as responsiveness. The broader textbook definition counts such a program as a simple agent, as the types below show; name the definition you use.

Step 4: Social ability: the agent interacts with other agents

Social ability is the ability to interact with other agents, and with people, when appropriate. Jennings and Wooldridge list it as the third part of flexibility: agents should be able to interact, when they deem it appropriate, with other agents and people. In practice that means asking a user for missing information, handing a subtask to another agent, or asking a person to approve a risky step, as covered in human in the loop AI agents.

Input: messages from other agents or people. Output: messages back, or a changed plan. The Encyclopedia of Life Support Systems (EOLSS) chapter on the weak notion of agent says that intelligence requires flexibility with respect to autonomous actions, so intelligent agents also exhibit responsive, social and pro-active behaviour alongside autonomy. You now have the full definition. Together the four steps are the loop below: the agent acts, the action changes the environment, and the agent senses the change.

1 · Autonomy2 · Reactivity3 · Proactiveness4 · Social abilityReplay

Step-by-step description

1. The agent acts, and the action changes the world.
2. It senses changes in its environment.
3. It pursues a goal, not only reactions.
4. It exchanges messages with other agents.

• An agent acts on its environment without step-by-step control.
• Each action changes the environment, and the agent senses those changes.
• An agent pursues a goal rather than only reactions.
• An agent exchanges messages with other agents.

In the figure, an agent acts on its environment without step-by-step control. Each action changes the environment, and the agent senses those changes. It pursues a goal rather than only reactions, and it exchanges messages with other agents.

What are the types of intelligent agents?

The standard textbook types of agent are simple reflex, model-based, goal-based, utility-based and learning agents. The IEEE Technology Navigator names simple reflex, model-based and goal-based agents among its types. A simple reflex agent maps the current percept straight to an action. A model-based agent also keeps an internal picture of the parts of the world it cannot see. A goal-based agent chooses actions by asking which will reach a goal, and a utility-based agent scores how good each outcome is.

A learning agent improves its own behaviour from experience. The textbook view of Russell and Norvig, in Artificial Intelligence: A Modern Approach, is broader than the four-property definition: anything that perceives through sensors and acts through actuators is an agent, and the simple reflex agent is the simplest type. Under the Wooldridge–Jennings definition that same program is an agent but not an intelligent one.

How are intelligent agents used in AI?

Intelligent agents are used wherever software must act toward a goal in a changing environment. A personal assistant that schedules meetings and sends emails on a user's behalf is the example in the critical analysis arXiv 2506.01463. The IEEE entry also covers multi-agent settings, where several agents interact; that is the social ability property. LLM-based agents, in Anthropic's sense, direct their own processes and tool use to finish open-ended tasks, and a customer-support agent is one such design. The oldest case is a control loop such as the toy room below, which senses a temperature and switches a heater.

Worked example

A toy room shows the difference between reacting and planning. Everything here is illustrative and computed in a short script.¹ The room starts at 15 degrees and the run has 16 time steps, numbered 0 to 15, and someone occupies the room only in steps 12 to 15.¹ Three rules govern the room. First, a heater adds 1.5 degrees per step when it is on.¹ Second, the room leaks heat: each step it loses 10% of the gap between its temperature and the outside temperature of 10 degrees, so an unheated room drifts toward 10.¹ Third, each agent decides, and comfort is checked, at the start of each step using the temperature at that moment.¹

The performance measure is the rule that scores the run: one point for each occupied step with the temperature between 19.5 and 22 degrees, minus 0.2 for each step the heater runs in an empty room.¹ Three agents try to maximise it. The reactive agent heats only when someone is present and the room is below 21 degrees. The always-heat agent heats whenever the room is below 21 degrees, ignoring who is there. The look-ahead agent knows the occupancy schedule. It assumes the heater nets about 1 degree per step, counts the steps needed to reach 21 degrees, and starts heating when that count reaches the steps left before someone arrives, which happens at step 4, eight steps early.¹

The reactive agent senses and responds, but every degree it adds arrives after the person does, so it scores zero comfortable steps.¹ The always-heat agent reaches comfort by brute force and wastes 11 empty-room heater steps.¹ The look-ahead agent uses its goal and the schedule, so it is comfortable for all four occupied steps with only 8 wasted ones.¹ That is proactiveness paying off in numbers. The temperatures at the start of selected steps show why the scores differ.

These temperatures come from the saved trace of the run.¹

Text description of the diagram

• In this illustrative run the look-ahead agent scores 2.4.
• The always-heat agent scores 1.8 and the reactive agent scores 0.0.

Trade-offs and edge cases

More agency costs more. Anthropic's guidance on building effective agents is to find the simplest solution possible and increase complexity only when needed. It adds that agentic systems often trade latency and cost for better task performance, and that for many applications a single LLM call with retrieval and in-context examples is usually enough.

Anthropic also draws a line inside the family. Workflows are systems where LLMs and tools are orchestrated through predefined code paths; agents are systems where LLMs dynamically direct their own processes and tool usage, keeping control over how they accomplish tasks.

Text description of the diagram

• Workflows run on predefined code paths.
• In agents the model directs its own steps and tools.

In the table, workflows run on predefined code paths, and in agents the model directs its own steps and tools. Cases to keep in mind:

• Two senses of "agent". Anthropic notes that the word covers several meanings, from fully autonomous systems to prescriptive workflows. State which one you mean.
• Strong and weak notions. The classic four-property definition is called the weak notion of agent in the EOLSS chapter, so a stronger notion exists in the literature; it usually adds mentalistic notions such as beliefs and intentions. Do not claim your definition is the only one.
• Loose vocabulary. A critical analysis (arXiv 2506.01463) argues that "agentic" is used loosely and mixes buzzwords with the established agent literature.
• Tool access. Tool protocols such as the Model Context Protocol (MCP) give an agent a shared way to reach tools; see what the Model Context Protocol is.
• Long tasks. Autonomy compounds errors; see why agents fail on long tasks.

How to answer it in an interview

The spoken structure: give the definition, name the four properties, split workflow from agent, then state the cost trade-off. For a full design answer, continue with designing an AI agent system.

Common mistakes

• "Any program that responds to input is an intelligent agent." (wrong) Under the Wooldridge–Jennings definition, responsiveness is one property of four, so such a program is an agent at most; the broader textbook definition would call it a simple reflex agent. "I say which definition I use: perceive-and-act for an agent, flexible autonomous action for an intelligent agent." (right)
• "An agent means a fully autonomous system." (wrong) The word has several meanings, including workflows. "I say whether the model or the code chooses the next step." (right)
• "Agents are always better than a single model call." (wrong) A single call with retrieval and examples is often enough. "I add agency only when the simpler design falls short." (right)
• "Intelligent agents are an LLM idea." (wrong) The definition predates LLMs by decades. "The four properties come from the 1990s; LLMs are one way to build them." (right)

Follow-up questions

What if heating an empty room cost more?

Each wasted step would subtract more from the score, so the gap between the always-heat and look-ahead agents would widen, because the first wastes 11 empty-room steps and the second 8.¹

What is the difference between an agent and a workflow?

In a workflow, code fixes the path through the model and tools. In an agent, the model decides its own steps and tool use, according to Anthropic (2024).

Is every chatbot an intelligent agent?

Not necessarily. Under the Jennings and Wooldridge (1998) definition, a chatbot that only answers each message is responsive but not proactive, and flexible behaviour needs more than responsiveness, so it is an agent at most, not an intelligent one. Under the broader textbook perceive-and-act definition it can count as a simple agent.

When should you not use an agent?

When a single call with retrieval and examples, or a fixed workflow, already does the job, because agentic systems often trade latency and cost for performance.

How do you tell whether an agent is any good?

Score it against a performance measure, as the toy room did. For real agents, check the outcome and the steps it took, which is the subject of agent evaluation.

Are intelligent agents new with large language models?

No. The four-property definition is credited to Wooldridge and Jennings in 1995 (IEEE Technology Navigator); LLMs are a newer way to build such systems.

Sources

1. Applications of Intelligent Agents (Jennings and Wooldridge), in Agent Technology: Foundations, Applications, and Markets — Springer, 1998 (copy in the UMBC agents archive), 1 Jan 1998
2. What Is an Intelligent Agent? — IEEE Technology Navigator (undated; read date shown), 26 Oct 2026
3. Weak notion of agent (sample chapter) — UNESCO Encyclopedia of Life Support Systems (undated; read date shown), 26 Oct 2026
4. A critical analysis of how agentic is used (arXiv 2506.01463, v1) — arXiv (v1, June 2025), 1 Jun 2025
5. Building effective agents — Anthropic Engineering, 19 Dec 2024