There are five classic types of AI agents, and they come from Russell and Norvig's textbook, Artificial Intelligence: A Modern Approach. They are simple reflex, model-based reflex, goal-based, utility-based and learning agents. Each adds one ability (rules, memory, a goal, a way to score outcomes, a way to improve) to the type below it.

Most agents sold in 2026 are built another way, around a language model that calls tools and sometimes hands work to other agents. This post covers both sets, with one real example per type. If you are new to the topic, start with what an AI agent is.

One precise point first. The book names four basic kinds of agent program and treats learning as a way any of them can improve. "Five types" is the common shorthand, and I explain the gap below.

Abstract grey, black and orange geometric shapes linked into one chain from left to right
Each type adds one ability to the one before it: rules, then memory, then a goal, then a score, then learning.

The five types at a glance

Each row keeps everything above it and adds one thing. For the sense-and-act loop all five share, see how AI agents work.

TypeWhat it addsThe question it asksReal exampleWhere it breaks
Simple reflex agentCondition-action rulesWhich rule matches what I see?A thermostat; a Gmail filterAnything it cannot see right now
Model-based reflex agentInternal state (memory)What is happening that I cannot see?A robot vacuum that maps the floorIt has no goal to plan toward
Goal-based agentA goal, plus search and planningWhat happens if I do this?Maps route search; a chess engineEvery route to the goal counts the same
Utility-based agentA function that scores outcomesHow good is each outcome?Google Maps eco-friendly routingOnly as good as its scoring
Learning agentA critic and a learning elementHow did that go, and what should change?AlphaGo ZeroLearns only from the feedback it gets

1. Simple reflex agents: if this, then that

A simple reflex agent looks only at its current percept (what its sensors report right now) and fires whichever condition-action rule matches. It keeps no memory and no model of the world.

A thermostat is the classic case. It switches the heat on below the set point and off above it, and never asks why the room got cold. In software, a Gmail filter rule (mail from this sender gets this label) works the same way.

It breaks on anything it cannot see right now. The textbook says a simple reflex agent works only when the right decision can be made from the current percept alone, which means a fully observable environment. Your filter has no idea the same sender wrote an hour ago to cancel.

2. Model-based reflex agents: rules plus memory

A model-based reflex agent keeps an internal state: a running picture of what the current percept does not show. It still acts by rules, now reading memory as well as sensors. Keeping that picture accurate needs two kinds of knowledge: how the world changes, and what the agent's own actions do.

Picture a robot vacuum that builds a map of the floor as it cleans. From the hallway it cannot see the kitchen, but the map says the kitchen is done, so it moves on.

The 2026 version is a support bot that remembers what the customer said three messages ago. The newest message only says "it still doesn't work"; the stored state knows which order it is about and which fix was already tried. That is short-term AI agent memory.

What it still lacks is a goal. It knows where it is. It has no idea where it is going.

3. Goal-based agents: planning toward a target

A goal-based agent holds a description of the state it wants and uses search and planning to reach it. It asks "what happens if I do this?" before it moves.

Route search in a maps app is the everyday example. The best route sometimes heads away from the destination briefly to reach it sooner. A rule like "always turn toward the destination" would never find it; a planner does, because it looks at the whole path. A chess engine searches moves and replies the same way.

Now tell an LLM agent "book a meeting with Priya next week". It plans the steps itself: check your calendar, propose times, send the invite once she picks one. Nobody wrote those steps as rules. That planning step is what agentic reasoning describes.

The weakness is easy to miss. A goal is binary. Reached or not reached. Any plan that gets there counts, even a slow or expensive one.

4. Utility-based agents: choosing the best outcome

A utility function fixes that. A goal says whether you arrived; utility says how good the arrival was. The agent scores each possible outcome and picks the best.

My favourite public example is Google Maps eco-friendly routing, announced on 30 March 2021 by Dane Glasgow, VP of Product for Google Maps. Maps would "default to the route with the lowest carbon footprint" when it has roughly the same ETA as the fastest route. If the greener route would noticeably lengthen the trip, Maps lets you "compare the relative CO2 impact between routes" and choose.

That is a utility function trading minutes against fuel. When the time cost is small, Maps picks the greener route; when it is big, Maps hands you the choice.

Compare it with section 3. Same product, two rungs. Plain route search is goal-based. Eco-friendly routing is utility-based.

In LLM agents, utility usually shows up as ranking options against a rubric. The score can only be as good as that rubric, so write it down.

5. Learning agents: getting better from feedback

A learning agent changes itself based on how things went. Russell and Norvig give it four components, which IBM's explainer also lists:

DeepMind's post on AlphaGo Zero, dated 18 October 2017, says it learned "simply by playing games against itself", starting "from completely random play" with no human game data. After three days of self-play, it beat the version that had beaten Lee Sedol "by 100 games to 0".

Claude Code's docs say it "builds auto memory as it works, saving learnings across sessions". That is learning by memory.

Why the textbook says four, not five

The book's own count is four, plus learning. On 9 October 2026 I read the University of Miami's CSC345 lecture slides (2026), which reproduce chapter 2. They quote the book: "We outline four basic kinds of agent program that embody the principles underlying almost all intelligent systems." Those are the first four rungs above.

Learning sits on top. The slides' summary reads: "All agents can be turned into learning agents that improve their performance through experience." "Five types" is fine shorthand for a blog. In an exam, I would write four agent programs plus learning.

The same summary adds: "The appropriate design of the agent program depends strongly on the nature of the task environment."

It lists six environment properties to check (fully or partially observable, deterministic or stochastic, episodic or sequential, static or dynamic, discrete or continuous, single-agent or multi-agent). A fully observable task is reflex territory. A partially observable one needs memory.

The three types you meet in 2026

The ladder describes how an agent decides; products in 2026 are described by how they are wired. Anthropic's "Building effective agents" (19 December 2024) has the clearest vocabulary I know.

Its building block is "an LLM enhanced with augmentations such as retrieval, tools, and memory". Then it draws one line: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage."

It names five workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer); see AI agent architecture patterns for more. I agree most with its advice: "finding the simplest solution possible, and only increasing complexity when needed".

Workflow agents

A workflow agent runs LLM steps inside a path you drew. The model handles the fuzzy step, such as drafting a reply; the code decides what happens next. AI agent vs workflow automation compares the two approaches.

In n8n, you attach a chat model and tools to the AI Agent node, and the docs are explicit: "You must connect at least one tool sub-node to an AI Agent node". The workflow around that node is still a path someone designed in advance.

Tool-using LLM agents

Here one model runs a loop and picks its own tools. Claude Code works this way; its docs call it "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools". It reaches outside tools through MCP, the Model Context Protocol.

That tool supply is growing fast. On the MCP adoption tracker on this site, the official MCP Registry listed 40,965 servers as of 9 October 2026, up 37% in 30 days. How an agent picks among them is the subject of AI agent tool use, explained.

Multi-agent systems

Several agents split one job, often with a lead agent handing pieces to subagents. In a 13 June 2025 engineering post on its Research feature, Anthropic reported that Claude Opus 4 leading Claude Sonnet 4 subagents "outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval".

The same post puts a price on it: "agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats". In its analysis, token usage alone explained 80% of the performance variance on BrowseComp.

I read the 15x as a price tag. Pay it for wide research across many sources; skip it when one agent can finish the job. For more, see multi-agent systems explained and single-agent vs multi-agent.

Mapping the 2026 types onto the ladder

2026 typeTextbook rungs it usesExample
Workflow agentReflex rules plus goal steps someone else plannedn8n AI Agent node inside a workflow
Tool-using LLM agentGoal-based planning, utility-style ranking, memory in place of learningClaude Code
Multi-agent systemGoal-based agents coordinated by a lead agentAnthropic's Research feature

If you build your own, check the framework's health too. On the framework leaderboard on this site, Dify ranked first of 43 open-source agent frameworks (95.8 of 100) as of 8 October 2026, ahead of n8n (93.1) and Mastra (84.9). Six of the 43, including AutoGen and MetaGPT, were archived or had no commit in 90 days.

Which type do you actually need?

Climb only as high as the task forces you. Here is how six common situations map to a type.

Two free tools help with the first call. The automation platform picker ranks Zapier, Make, n8n and 13 other platforms for a task you describe, with published prices. The Can AI do this task tool shows how much of a task you describe an agent can take on today.

Where Gravity fits

Gravity's agents are tool-using LLM agents that independent builders have already built and published. You describe a recurring task in plain words and pick one, so you never classify anything or build anything. Gravity is in public beta: Free is $0 with no card, and Mini Pilot is $15 a month for one automated agent. Explore agents.

Frequently asked questions

What are the 5 types of AI agents according to Russell and Norvig?

Russell and Norvig's textbook, Artificial Intelligence: A Modern Approach, describes simple reflex, model-based reflex, goal-based, utility-based and learning agents. Each type adds one ability to the type before it. Strictly, the book calls the first four basic agent programs and treats learning as a layer any of them can gain.

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

A simple reflex agent acts only on what it perceives right now, so it works well only when the environment is fully observable. A model-based reflex agent also keeps an internal state that tracks what it cannot currently see, such as earlier messages in a chat. Neither has a goal.

How does a goal-based agent differ from a utility-based agent?

A goal-based agent only knows whether an outcome reaches its goal, so any plan that arrives counts. A utility-based agent scores how good each outcome is and picks the best. Google Maps shows both: route search is goal-based, while eco-friendly routing (March 2021) trades travel time against fuel.

Which four components make up a learning agent?

Russell and Norvig name four: the performance element chooses actions, the critic judges results against a standard and gives feedback, the learning element changes the agent based on that feedback, and the problem generator suggests exploratory actions. AlphaGo Zero is a well-known example.

What type of AI agent is ChatGPT or Claude Code?

Both are best described as tool-using LLM agents: a model that runs a loop and picks its own tools. On the textbook ladder they act mostly like goal-based agents, with utility-style ranking and memory standing in for learning. The companion post, Is ChatGPT an AI Agent?, gives the longer answer.

Are multi-agent systems a sixth type of AI agent?

The textbook does not count them as one; Russell and Norvig treat single-agent versus multi-agent as a property of the task environment. In 2026 practice, a multi-agent system is an architecture where several agents split one job. Anthropic reported in June 2025 that these systems use about 15x more tokens than chats.

Sources