Ask "how many types of AI agents are there" and you will get a tidy answer: five. The standard taxonomy, set out in Russell and Norvig's textbook Artificial Intelligence: A Modern Approach, sorts agents into simple reflex, model-based reflex, goal-based, utility-based, and learning agents. The ladder runs from agents that just react to agents that plan, weigh trade-offs, and improve themselves. It is a clean way to see why some agents feel dumb and others feel almost thoughtful.

That textbook split is the right starting point, but it is not how builders talk day to day. So this guide does two things: it translates the five classic types into plain English with a real example for each, then maps them onto the practical categories you will meet when you pick an agent. For the ground floor, see what is an AI agent and agentic AI explained without jargon.

The classic taxonomy at a glance

The five-type taxonomy comes from Russell and Norvig's Artificial Intelligence: A Modern Approach, the standard university text used in AI courses worldwide. It arranges agents on a rising ladder of capability: each type adds one thing the one below it lacks. Simple reflex agents react. Model-based agents remember. Goal-based agents plan. Utility-based agents compare. Learning agents adapt. That single idea explains the whole list.

Here is the through-line. A simple reflex agent has only rules. Add an internal picture of the world and you get a model-based agent. Give that agent a target and it becomes goal-based. Let it score options against each other and it is utility-based. Finally, let it change its own behaviour from feedback, and you have a learning agent. Each rung sits on top of the last.

One thing worth flagging early. These are not five competing products you buy. They are levels of sophistication, and a single real system often spans several rungs at once. The taxonomy is a lens for understanding behaviour, not a shopping menu. Keep that in mind as we walk each type with a concrete example.

1. Simple reflex agents

A simple reflex agent maps the current input straight to an action using fixed condition-action rules, with no memory and no model of the world. In Russell and Norvig's framing, it is the most basic agent: "if condition, then action," nothing more. It cannot reason about the past or the future. It only sees now, matches a rule, and acts. That is its entire mind.

Real example: a thermostat

A classic thermostat is the textbook simple reflex agent. The rule is plain: if the temperature drops below the set point, switch on the heat; if it rises above, switch it off. It does not remember yesterday or predict tomorrow. A spam filter built only on keyword rules works the same way: see a banned word, bin the message. Fast, cheap, and brittle the moment the world stops matching its rules.

The limit is obvious. A simple reflex agent fails the instant it needs to know something it cannot currently see. If the temperature sensor is missing, the thermostat is blind. That gap is exactly what the next type fixes. To see where reflex behaviour sits inside a full agent, the breakdown in how AI agents work is the natural companion read.

2. Model-based reflex agents

A model-based reflex agent keeps an internal model of how the world works, so it can act on things it cannot directly see right now. Russell and Norvig describe it as a reflex agent that maintains internal state, updated from each new input plus its knowledge of how the world changes. It still uses rules, but those rules now read from memory, not just the current input.

Real example: a robot vacuum

A robot vacuum is a tidy example. It builds a map of the room, remembers where it has already cleaned, and tracks where walls and furniture sit even when its sensors point elsewhere. A self-driving car's lane-keeping does the same, holding a model of nearby vehicles it cannot see this exact frame. The internal model is what lets these agents behave sensibly in a world they only glimpse one slice at a time.

This rung matters because the real world is partly hidden. Sensors are noisy and limited, so an agent that only reacts to the visible moment keeps tripping over what it forgot. Holding state is the first big step toward an agent that seems to understand its surroundings. Still, it reacts; it does not yet aim at anything.

3. Goal-based agents

A goal-based agent holds a target and chooses actions by asking which ones move it closer to that goal, planning ahead rather than just reacting. Russell and Norvig note that goals let an agent consider the future: "what will happen if I do this?" instead of "what rule fires now?" This is the rung where search and planning enter, and where an agent can pick a worse-looking move that leads somewhere better.

Real example: a route planner

A GPS navigation app is a clean goal-based agent. The goal is your destination. It searches across possible routes, including ones that briefly head the "wrong" way, and picks the path that actually reaches the target. A chess engine looking for checkmate behaves the same way: it plans sequences toward a defined end state, not a one-step reaction. The goal reshapes every choice along the route.

The jump from reflex to goal-based is the jump from reacting to reasoning. Most of the "agentic" excitement in 2026 is really excitement about goal-based behaviour finally working at scale; the language model supplies the planning a reflex rule never could. If you want the planning side in depth, AI agent architecture patterns explained covers the common structures.

4. Utility-based agents

A utility-based agent goes past a single goal: it scores possible outcomes on a utility function and picks the option with the highest expected value, so it can balance competing goals. Russell and Norvig point out that goals only tell you "reached or not"; utility tells you "how good," which matters when there are many ways to succeed and they are not equally good. It chooses the best, not merely the acceptable.

Real example: a ride dispatch system

A ride-hailing dispatch engine is a strong utility-based example. Several drivers could take a trip, so the system scores each pairing on driver distance, wait time, fuel, and fairness, then picks the assignment with the best overall value. A logistics planner weighing speed against cost against emissions does the same. When goals conflict, utility is the tiebreaker that turns "any valid answer" into "the best answer."

Why does this rung matter so much in practice? Because real work is full of trade-offs. In our experience reviewing agent builds, the ones that feel "smart" to users are almost always the ones doing implicit utility scoring: ranking three good draft replies and sending the best, not the first. That judgment is what separates a helpful agent from a merely functional one.

5. Learning agents

A learning agent improves its own behaviour from experience, so it gets better over time instead of staying fixed. Russell and Norvig split it into four parts: a performance element that acts, a critic that judges results against feedback, a learning element that updates the agent, and a problem generator that suggests new things to try. The other four types can all be made into learning agents by bolting this loop on top.

Real example: a recommendation feed

A streaming recommendation feed is the everyday learning agent. It suggests titles, watches whether you click and finish them, and quietly adjusts what it shows next. Spam filters that adapt to new tricks and fraud-detection models that retrain on fresh cases work the same way. Each run feeds the next; the agent you use tomorrow is shaped by everyone's behaviour today.

Learning is the top rung because it makes every other type better without a human rewriting rules. The contrarian point worth making: most production agents today learn far less than people assume. The model's weights are frozen at use time, and "learning" usually means improved prompts, tools, or retrieved examples around a fixed model, not the agent rewiring itself mid-task.

The modern practical split

The textbook five describe how an agent thinks; builders also sort agents by how they are used. Anthropic's engineering guide "Building Effective Agents" (2024) draws a practical line between workflows, where steps are fixed in code, and agents, where a model directs its own actions. That split, plus a few others below, is closer to the choices you actually face when picking an agent.

Assistive versus autonomous

An assistive agent stays in the loop with you: it drafts, suggests, and waits for your go-ahead, like a coding copilot proposing the next line. An autonomous agent runs a whole task end to end and reports back when it is done. The line is who holds the steering wheel. Most useful real-world agents sit somewhere in between, autonomous on the routine parts, assistive at the risky decisions.

Single-agent versus multi-agent

A single-agent system has one agent doing the whole job. A multi-agent system splits work across several agents that hand off to each other, like a researcher feeding a writer feeding an editor. Multi-agent setups shine on big tasks with distinct sub-jobs, but they add coordination cost. How agents combine into larger systems is the subject of AI agent composability explained.

Tool-using language model agents

The agent type most people mean in 2026 is the tool-using language model agent. It reads a goal in plain language, plans steps, calls tools such as search or a calendar, checks the results, and adjusts. It quietly blends goal-based planning, utility-style ranking, and a learning loop around a fixed model. To see how this differs from older chatbots, read AI agent vs chatbot vs assistant.

What you actually deploy

Here is the honest summary: nobody ships a "pure" textbook type. Anthropic's "Building Effective Agents" (2024) makes the case plainly, that the best results come from composing simple, well-understood patterns rather than chasing one grand architecture. The agent you run is a hybrid, goal-based at heart, utility-aware in its choices, wrapped in a thin learning loop, and using tools to touch the real world.

On a platform like Gravity, you never have to classify any of this. You describe the outcome you want in plain words, and an expert-built agent, already designed with the right mix of planning, ranking, and tool use, runs the task and hands back the result in about 60 seconds. We have found users rarely care which textbook rung an agent sits on; they care that it finished the job and that the bill matches the work done.

That said, knowing the taxonomy still pays off. It helps you judge whether an agent is built for the job: a task with real trade-offs needs utility-style ranking, a task that should improve over time needs a learning loop, and a one-shot reaction may need nothing fancy. For how that maps to running costs, see AI agent cost models explained, and for how vendors prove capability, AI agent benchmarks explained.

Frequently asked questions

What are the main types of AI agents?

The classic taxonomy names five types: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. They are ordered by how much each one knows and plans. Modern practice adds a second split by autonomy and structure: assistive versus autonomous, and single-agent versus multi-agent systems.

What are the 5 types of AI agents?

The five types are simple reflex agents that react to current input, model-based reflex agents that track hidden state, goal-based agents that plan toward a target, utility-based agents that weigh trade-offs to pick the best option, and learning agents that improve from feedback over time. This ordering comes from Russell and Norvig.

What is the difference between a reflex agent and a goal-based agent?

A reflex agent maps the current input straight to an action using fixed rules; it never considers the future. A goal-based agent holds a target and plans a sequence of actions to reach it, so it can pick moves that look worse now but lead toward the goal. Goal-based agents think ahead; reflex agents do not.

What is a learning agent?

A learning agent improves its own behaviour from experience instead of staying fixed. It has a performance element that acts, a critic that judges results against feedback, and a learning element that adjusts the agent so the next run is better. Recommendation systems and spam filters are everyday examples.

Which type of AI agent is most useful in practice?

For real work, the most useful agents blend goal-based planning with tool use, and often learning on top. A tool-using language model agent reads a goal, plans steps, calls tools, and checks results. That hybrid, not a pure textbook category, is what people deploy to finish actual tasks today.

Three takeaways before you close this tab

Sources