All posts
AI Employee#types of AI agents#types of agents in AI#simple reflex agent

Types of AI agents: the textbook five, and how a chatbot grew into a coworker

The five types of AI agents from the textbook (simple reflex, model-based reflex, goal-based, utility-based and learning), explained with everyday examples, then the types you will actually meet at work: rule-based chatbots, conversational AI, assistants, tool-using agents, multi-agent systems and the AI coworker.

Md Shohel· October 26, 2026· 16 min read
Listen to this post0:00 / 18:10
AI narration of this post, read in full

Search for "types of AI agents" and you get two kinds of answer that barely overlap.

The first comes from computer science courses. It lists five types, from the simple reflex agent up to the learning agent, usually with a thermostat and a self-driving car as the examples. It's correct, and it was written for students taking an exam.

The second comes from software vendors. It lists whatever the vendor sells, plus a few things the vendor doesn't, arranged so its own product sits at the top.

If you run a business and you're trying to work out what the thing in front of you actually is, neither list helps much on its own. So this post does both. First the five textbook types of agents in AI, in plain language, with an office example for each. Then the short history of how conversational AI went from a 1960s script that imitated a therapist to software that holds a job, because that history is where the types you'll meet at work came from. We make Sanaf, an AI employee, so we have an interest in the last section. The rest stands on its own.

The short answer

An AI agent is software that perceives something about its situation, decides what to do, and acts. The classic textbook sorts agents into five types by how they decide:

  1. Simple reflex agents act on what they see right now, using fixed if-then rules.
  2. Model-based reflex agents keep a memory of what they can't see and use rules on top of it.
  3. Goal-based agents pick actions by asking which one moves them closer to a goal.
  4. Utility-based agents weigh several goals and pick the action with the best overall outcome.
  5. Learning agents improve their own behaviour from feedback over time.

In business software you'll more often hear a second set of names: rule-based chatbots, conversational AI, AI assistants, tool-using AI agents, multi-agent systems and AI coworkers or AI employees. Those are product categories rather than textbook types, and each one maps onto the five above.

The five textbook types of AI agents

The five types come from Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig, first published in 1995 and still the standard university text. Most "what are the 5 types of agents" answers online are summaries of its second chapter. The book defines an agent broadly: anything that perceives its environment through sensors and acts on it through actuators. A person is an agent. So is a thermostat.

What separates the types is how much the agent knows and how it chooses.

TypeHow it decidesTextbook exampleOffice example
Simple reflexIf this, then that, on the current input onlyThermostatEmail filter that moves anything from a domain to a folder
Model-based reflexRules plus a memory of what it can't currently seeCar that remembers a vehicle in its blind spotTicket system that knows a customer already wrote twice this week
Goal-basedWhich action gets me closer to the goal?Route plannerScheduler finding a slot that works for four people
Utility-basedWhich action gives the best outcome across several goals?Self-driving car trading speed against safetyChoosing which overdue invoice to chase first
LearningChanges its own rules from feedbackSpam filter that improves when you mark mailAssistant that stops drafting replies you always rewrite

1. Simple reflex agents

A simple reflex agent looks at the current input, matches it against a rule and acts. It has no memory and no idea of a goal. The thermostat sees 18 degrees, the rule says "below 20, heat on", and the heat comes on.

Most office automation is this type. A mail rule that files every message from a supplier's domain, a form that sends a fixed confirmation, a Zapier zap that posts to a channel when a row is added. They're fast, cheap and predictable, and they break the moment the input doesn't match what the rule expected. The supplier changes its email domain and the filter quietly stops working.

2. Model-based reflex agents

A model-based reflex agent keeps an internal picture of the parts of the world it can't see directly, and updates it as things happen. It still acts on rules, but the rules can refer to that picture.

The textbook example is a car that remembers there's a vehicle in its blind spot even after the mirror stops showing it. The office version is a helpdesk that knows this customer already opened two tickets about the same order, so the third one goes straight to a person instead of getting the standard reply. The rule is still a rule. It just has memory to work with.

3. Goal-based agents

A goal-based agent is given an outcome and works out which actions lead there. It searches or plans instead of reacting. The question it asks is "if I do this, will I be closer to the goal?"

Route planners are the classic case. Scheduling is the everyday one. "Find 45 minutes next week when all four of us are free, and avoid Friday afternoon" can't be solved by one if-then rule. The agent has to look at calendars, try options and keep the one that meets the goal.

4. Utility-based agents

A goal-based agent only knows whether a goal is met. A utility-based agent can tell how well it's met, which matters when goals compete. It scores possible outcomes and picks the best one overall.

A self-driving car balances arriving quickly against arriving safely. A finance team balances chasing the largest overdue invoice against chasing the one most likely to pay this week, and against not annoying a client who is already unhappy. Most real business decisions look like this, with several things that matter and no option that wins on all of them.

5. Learning agents

A learning agent has a part that does the work and a part that criticises the work, and it uses the criticism to change how it behaves next time. Any of the four types above can also learn.

Your spam filter is a learning agent. Every time you mark a message as spam or rescue one from the folder, it adjusts. At work, the useful version is an assistant that notices you rewrite every draft it gives you for a particular client and starts writing them closer to how you would.

Two more you'll see on lists

Some lists add hierarchical agents, where a top-level agent splits a task and hands the parts to lower-level agents, and multi-agent systems, where several agents work on a shared problem and either cooperate or compete. These aren't separate decision types in the textbook sense. They describe how agents are organised rather than how each one thinks. They matter a lot in modern products, so they come up again below.

From chatbot to coworker: how conversational AI got here

The textbook types describe how an agent decides. The products people actually buy grew out of a different line, the attempt to make software you can talk to. Following that line explains why "chatbot", "assistant", "agent" and "AI employee" are four different things, even though vendors use the words loosely.

1966: ELIZA, the first chatbot

The history of chatbots usually starts with ELIZA, written by Joseph Weizenbaum at MIT and described in a paper in 1966. Its best known script, DOCTOR, imitated a therapist by turning your sentences back into questions. Type "I'm worried about work" and it might answer "Why are you worried about work?"

ELIZA understood nothing. It matched keywords and rearranged your words with templates. It was a simple reflex agent with a conversational surface. People still confided in it, and Weizenbaum later wrote about his discomfort that his own secretary asked him to leave the room so she could talk to it privately. That reaction, people treating fluent text as understanding, has followed every chatbot since.

The rule-based chatbot years

For the next few decades almost every chatbot was rule based. A rule-based chatbot follows a decision tree that a person wrote: show a menu, match a few keywords, follow the branch, hand off to a human when the tree runs out. When messaging platforms opened up to bots in the mid 2010s, businesses built thousands of them for order tracking, store hours and FAQs.

They worked for narrow questions and failed in a familiar way. Ask anything the tree didn't anticipate and you got "Sorry, I didn't understand that" until you gave up or typed "agent" enough times. Those bots were simple reflex agents, sometimes with a little memory, which made them model-based reflex agents.

Voice assistants: the first assistants most people met

Siri arrived on the iPhone in 2011 and Alexa on the Echo in 2014. They added speech recognition and a large set of built-in skills, like timers, weather, music and smart home controls. Under the surface they were still mostly intent matching: figure out which of a known set of things you want, fill in the blanks, run the matching function.

They also taught people the main limit of an assistant. It waits to be asked and does one thing per request. Ask it to set a reminder and it sets the reminder. It doesn't notice a week later that you never did the thing.

Conversational AI and large language models

Conversational AI is the umbrella term for software that holds a natural conversation, understanding free text or speech and replying in kind, rather than walking you down a menu. For years the phrase covered careful intent-matching systems. Large language models changed what it meant.

When ChatGPT launched on 30 November 2022, the reply was no longer chosen from a list. The model generated it. It could handle questions nobody had planned for, keep context across a long conversation and write in any register. That's the difference behind the "chatbot vs conversational AI" question: a rule-based chatbot follows a script, while conversational AI understands and produces open language. In practice nearly every chatbot launched now is built on a language model, so the two words are converging.

What the early language model chatbots still couldn't do was act. They could write the email. They couldn't send it.

Tool use: when the chatbot got hands

The next step was letting the model call tools. Instead of only producing text, the model can produce a structured request, such as "search the calendar for Thursday" or "create a draft in Gmail to this address", and the surrounding software runs it and returns the result. The model reads the result and decides what to do next.

This is the point where a chatbot becomes an AI agent in the modern sense. It's a goal-based agent, sometimes a utility-based one, whose reasoning is done by a language model and whose actions go through tools. Open standards for connecting tools, such as the Model Context Protocol that Anthropic published in late 2024, made it much easier to plug the same agent into many apps. If you want the longer version of that definition, is ChatGPT an AI agent? walks through it, and what is agentic AI? covers the loop that makes software agentic.

Multi-agent systems

Once one agent could use tools, people started putting several together. In a multi-agent system, each agent has a narrower role, such as one that researches, one that drafts and one that checks, and they pass work between them. Often one agent coordinates and the others report to it, which is the hierarchical pattern from the textbook list.

The benefit is focus: an agent with one job and a short set of instructions is easier to keep on track than one agent asked to do everything. The cost is coordination. Somebody has to decide who does what, and the hand-offs are where things get lost, the same as with people.

The AI coworker

The last step is less about intelligence and more about where the agent lives and what it's responsible for. An AI coworker, or AI employee, is an agent given an ongoing job instead of a single request. It sits in the team's chat, works inside the company's own apps, runs on a schedule as well as on request, remembers what happened last week, asks before doing anything it can't undo, and keeps a record people can check.

That changes the questions a buyer asks. With a chatbot you ask whether the answers are good. With an AI coworker you ask what it's allowed to touch, who approves what, and what happens when it's wrong. Those are the same questions you'd ask about a new hire, which is the point of the name. The AI coworker page goes further into what that looks like day to day.

The types you'll actually meet at work

Put the textbook and the history together and you get a practical list. Each line here is a product category, with the textbook type underneath it.

What it's calledWhat it doesTextbook type underneath
Rule-based automationRuns fixed steps when a trigger firesSimple reflex
Rule-based chatbotWalks people through a decision treeSimple or model-based reflex
Conversational AI chatbotAnswers open questions in natural languageModel-based, generating replies with a language model
AI assistantDoes one task when asked, with some toolsGoal-based, waits for a request
AI agentPlans steps and uses tools to reach a goalGoal-based or utility-based, often learning
Multi-agent systemSeveral agents split a larger taskHierarchical or cooperative agents
AI coworker or AI employeeHolds an ongoing job across your apps, with approvals and a recordA coordinating agent with agents under it

Two things are worth knowing about that table.

The rows aren't better or worse versions of each other. A simple reflex agent is exactly right for "file every invoice email into the Invoices folder". Using a language model agent for that job adds cost and a small chance of a creative mistake for no gain. AI agent vs chatbot, assistant and automation makes this case in more detail with two questions that sort any product.

The labels on products don't follow the table. Plenty of tools sold as "AI agents" are rule-based automations with a chat box, and some conversational tools quietly take real actions. The way to tell is to ask what it does when nobody is watching and what it does when the input surprises it.

How to tell which type you're looking at

Five questions sort almost any product into the right row:

  1. Does it act, or only answer? If all it produces is text for you to copy, it's a chatbot or an assistant, however clever.
  2. Does it act when nobody asks? If it only responds to a message, it's an assistant. If it runs on a schedule or on events, like a new email or a form submission, it's closer to an agent.
  3. What happens when the input doesn't match? A rule-based tool fails or does nothing. A language model agent improvises, which is useful and also the reason you want approvals.
  4. What does it remember? No memory means reflex. Memory of the conversation means model based. Memory of your business, your clients and last month's decisions is what a coworker needs.
  5. Who approves what? A serious agent has a clear line between what it may do alone and what it must ask about. If a vendor can't tell you where that line is, assume there isn't one.

Where Sanaf fits

Sanaf is an AI employee. In the terms of this post, it's a coordinating agent with any number of agents reporting to it. Sanaf is the Chief of Staff, and you can add agents for particular jobs, which what is an AI chief of staff? explains. It lives in your team's Slack or Microsoft Teams and works inside the apps you connect, such as Gmail, Google Calendar and HubSpot, and only those.

You give it jobs from a roster, such as inbox, lead follow-up, booking or invoicing; the AI employee roster lists them all. It runs them on request and on a schedule. It asks before anything it can't undo, and each job leaves a work log the team can read. The security page covers how access and approvals work.

On price, you start with $120 of free credit and no card, and that free credit lasts 30 days. After that you buy credits from $20 when you need them, and credit you buy never expires. You pay for the work, not per seat. The pricing page has the details, and hiring an AI employee shows what the first day looks like.

A fair test of any of the types above, ours included, is one job you already do every week. Give it that job for a fortnight with approvals on, read what it wanted to do, and decide from the record rather than the demo.

Questions people ask about types of AI agents

What are the types of AI agents?

The classic types of AI agents are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents. Lists often add hierarchical agents and multi-agent systems, which describe how several agents are organised rather than how one agent decides.

What are the 5 types of agents in AI?

The five types of agents in AI, from Russell and Norvig's textbook, are: simple reflex agents, which act on current input with if-then rules; model-based reflex agents, which add a memory of what they can't see; goal-based agents, which plan toward a goal; utility-based agents, which pick the best outcome among competing goals; and learning agents, which improve from feedback.

What is a simple reflex agent?

A simple reflex agent acts only on its current input using fixed condition-action rules, with no memory and no goal. A thermostat and an email filter rule are simple reflex agents.

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

A goal-based agent only knows whether a goal is reached. A utility-based agent scores how good each outcome is, so it can choose well when goals compete, such as speed against safety or cost against quality.

What is a multi-agent system?

A multi-agent system is a setup where several AI agents work on a shared task, each with its own role, passing work between them. Often one agent coordinates and the others report to it.

What is conversational AI?

Conversational AI is software that understands and replies in natural language, by text or voice, instead of following a fixed menu. Modern conversational AI is built on large language models, which generate replies rather than choosing them from a script.

What is the difference between a chatbot and conversational AI?

A traditional chatbot follows a script or decision tree written by a person and fails on questions outside it. Conversational AI understands open language and generates its own replies. Most new chatbots are now built on conversational AI, so the terms increasingly overlap.

What was the first chatbot?

ELIZA, written by Joseph Weizenbaum at MIT and described in 1966, is generally called the first chatbot. It imitated a therapist by matching keywords and turning the user's sentences into questions.

What are the types of chatbots?

The main types of chatbots are rule-based chatbots, which follow a decision tree; keyword or intent-matching chatbots, which map messages to known requests; and AI chatbots built on language models, which generate replies. AI chatbots that can also call tools and take actions are usually called AI agents.

Is an AI employee a type of AI agent?

Yes. An AI employee, or AI coworker, is an AI agent given an ongoing job instead of single requests. It works in a team's chat and apps, runs on schedules and events, remembers past work, asks before actions it can't undo and keeps a record people can review.

Get the next one by email

We write when we have learned something worth passing on, so nothing arrives unless it is worth your time.

Unsubscribe in one click, any time. Privacy policy