What is an intelligent agent in AI? The five types

Ask three software vendors what an AI agent is and you will get three answers, and each one will describe the vendor's own product. The older, plainer meaning comes from the textbook that most university AI courses still use. It is worth knowing, because it tells you what to ask before you let any software act on your behalf.

This guide covers that meaning, the five classic types of agent with an example of each, and where today's language-model agents fit. It ends with what an agent can get wrong and what Gonzo's own agents are and are not.

What an intelligent agent is

An intelligent agent is a system that perceives its environment and acts on it to reach a goal. That is the framing in Chapter 2 of Russell and Norvig's Artificial Intelligence: A Modern Approach, whose table of contents covers agents and environments, performance measures and rationality, and the structure of agents.^[1]

The idea reduces to a loop with three parts:

• Percepts. What the agent can sense. For a thermostat, a temperature. For a software agent, an API response, a webhook, a database row or an incoming message.
• Decision. How it chooses what to do next, given what it senses and what it knows.
• Actions. What it can change. A thermostat switches a heater. A software agent sends a message, updates a CRM record or opens a ticket.

A rational agent, in the textbook's sense, is one that picks the action expected to do best against a performance measure, given what it has perceived so far. Two words in that sentence carry the weight. "Expected" means the agent can be wrong in a given case and still be well designed. "Performance measure" means someone has to define what counts as doing well, and that definition is where many business agents go wrong.

The five types, from simplest to most capable

The textbook's section on the structure of agents lists five designs: simple reflex, model-based reflex, goal-based, utility-based and learning agents.^[1] Each one adds a capability to the one before it. They are design patterns, not products, and a real system often mixes two or three.

Simple reflex agents

A simple reflex agent maps what it senses right now to an action, with no memory. If a condition holds, it does the thing.

A thermostat is the standard example. A software version is a rule such as "if a new form submission arrives, send the confirmation email". Such an agent is fast, cheap and easy to test. It fails when the right action depends on something it cannot see in the current moment, such as whether this person already got a confirmation yesterday.

Model-based reflex agents

A model-based reflex agent keeps an internal picture of the world, updated from what it senses. It still follows rules, but the rules can refer to history.

A login check that remembers which devices and places you normally sign in from is one example. Because it holds state, it can tell a new device from a familiar one. A reply system that remembers a customer already asked about delivery is another. Keeping state is also where bugs enter: if the stored picture drifts from reality, every later decision inherits the error.

Goal-based agents

A goal-based agent is given a target and searches or plans for a sequence of actions that reaches it. A route planner is the classic case: the goal is the destination, and the agent considers possible paths. A booking agent has the same shape. It wants an appointment inside a window and tries slots until one fits.

You may see "knowledge-based" or "logic" agents listed as a fifth type in articles on this subject, including an earlier version of this one. In the textbook, knowledge-based agents belong to a separate chapter on logic. They describe how an agent can represent and reason over facts, not a sixth position on the ladder above.^[1]

Utility-based agents

A utility-based agent scores outcomes and picks the action with the best expected score. This matters when goals conflict or results are uncertain, because "reaches the goal" stops being enough. A delivery route that arrives 10 minutes sooner but costs more is not obviously better, and a utility function is how the agent weighs the two.

The weak point is the score itself. An agent optimizes the number it is given. If the number is clicks, you will get clicks, whether or not they pay.

Learning agents

A learning agent improves from experience. It watches the results of its actions and changes how it decides. A spam filter that updates from messages people mark as spam is a familiar one. So is a system that shifts traffic among several versions of a headline based on which one people respond to.

Learning buys adaptability and costs predictability. A learned rule can be hard to explain afterward, so a learning agent needs a record of what it did and why.

The five at a glance

Where language-model agents fit

Most of the agents sold today put a large language model in the decision step. The five types still apply, because they describe how a decision is organized, not what technology makes it.

Anthropic draws a useful line here. In its engineering write-up on building agents, it calls systems where models and tools run through predefined code paths "workflows", and systems where the model directs its own process and tool use "agents".^[2] Its advice is to find the simplest solution that works and add complexity only when needed, which may mean not using an agent at all. The same page says autonomy brings higher costs and the potential for compounding errors, and recommends extensive testing in sandboxed environments along with guardrails.^[2] The post is dated December 2024, and Anthropic's own note says much of the tooling it describes has changed since. We cite it for the distinction, not the tooling.

In practice that gives you a rule of thumb. If the steps are known in advance, write them down as a workflow. Reach for an agent that chooses its own steps only when the path cannot be known beforehand.

What can go wrong

An agent that acts for you can fail in ways a report cannot. Four failures are worth planning for.

It optimizes the wrong thing. A utility-based or learning agent pursues the number it is given. Pick a measure that tracks what you want, and check it against real outcomes, not only the metric.

You cannot see why it did something. The more an agent learns or improvises, the less an explanation comes for free. Log every action and the input that led to it, so a person can reconstruct a decision.

It has more access than the job needs. An agent that can read a database and send email can also read the wrong rows and send the wrong email. Give it the narrowest access that does the work, and treat it like a new employee on day one.

Errors compound. If each step builds on the last, an early mistake travels. Test in a sandbox with fake data first, start with low-stakes tasks, and keep a person in the loop for anything involving money or a complaint.

Where agents are used outside business

Not every agent runs a company process. In games and robotics, developers train agents inside a simulation first, where a mistake costs nothing. Unity's open-source ML-Agents toolkit, for example, lets games and simulations serve as environments for training agents with reinforcement learning and imitation learning.^[3] The same logic carries over to business use: rehearse on fake data before the real thing.

What Gonzo's agents are, and are not

Gonzo Digital builds AI agents in Python, not canvases of connected apps. All nine of our systems are agents of that kind. They are also, in the textbook's terms, mostly plain designs. Rules decide what gets said and sent, a stored conversation state supplies the memory, and a language model is used only to word a message. That makes them close to model-based reflex agents with a goal attached. This is our own classification, not a standard label.

What they do not do matters more. Each message and action passes the same seven checks in the same order, from a kill switch to an audit log, and every system starts in a dry run where nothing is sent until a person has read the outbox. The search and AI-answer agents propose changes and have no way to apply them. A person decides. You can read the full list of checks on how it works.

We should also be exact about status. The systems have run only in a sandbox, with simulated customers. None has run against a live mailbox or payment account, and Gonzo has no paying clients, so there is no client result to point to. If you want one of these agents to answer new inquiries, the Lead Response Engine is the one that starts there. If you are deciding between a flow tool and an agent for a single task, the comparison with building it yourself says where the flow tool wins.