There are five classic types of AI agents: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. That’s the textbook answer to “what are the types of agents in AI,” and it’s correct. But it won’t help you decide what to actually build. A more useful way to think about types of AI agents is by how much freedom you give them: assistants, workflow agents, and autonomous agents. The textbook tells you what agents can do. The autonomy sort tells you what to build.
The 5 classic types of AI agents
These five come from Russell and Norvig’s Artificial Intelligence: A Modern Approach, the standard textbook since 1995. They show up in every college AI course. Here’s what each one does, in plain language.
1. Simple reflex agents act on what they see right now. No memory, no planning. A thermostat is the perfect example: the room is cold, so it turns on. It doesn’t remember yesterday’s temperature or think about tomorrow’s.
2. Model-based reflex agents keep a picture of the world in their head. They track things they can’t see right now. Think of a GPS that knows there’s a traffic jam two miles ahead even though you can’t see it yet.
3. Goal-based agents have an objective and plan steps to reach it. A chess AI doesn’t just react to the current board. It plans three moves ahead because it has a goal: win the game.
4. Utility-based agents pick the best action, not just one that works. Imagine a delivery app routing five drivers. Any route gets the packages there. The utility agent finds the cheapest, fastest combination.
5. Learning agents get better over time. They watch what works, update their approach, and improve. Your email spam filter learns from the junk you flag. After a few weeks, it catches things it missed on day one.
| Type | One-line description | Everyday example |
|---|---|---|
| Simple reflex | Sees → acts, no memory | Thermostat |
| Model-based reflex | Tracks what it can’t see | GPS with traffic data |
| Goal-based | Plans steps toward a goal | Chess AI |
| Utility-based | Picks the best option of many | Delivery route optimizer |
| Learning | Improves with experience | Spam filter |
These five are useful for understanding how AI thinks. But if you’re trying to figure out what kind of agent to build or buy for your business, they don’t help much. A spam filter and a self-driving car are both “learning agents,” but the gap between them is massive. You need a different sort.
A more useful sort: agents by autonomy level
A recent paper from Feng, McDonald and Zhang (2026) proposed five “levels of autonomy” for AI agents, defined not by what the agent can do, but by what you (the human) do. Your role shifts from operator to observer as autonomy goes up.
Anthropic (the company behind Claude) draws a similar line. They split AI systems into “workflows” (you design the steps, the AI runs them) and “agents” (the AI decides its own steps). That distinction is more helpful than the textbook five, because it tells you what to actually expect.
Here’s the practical sort. Three categories, from least to most freedom.
Assistants: you drive, it helps
You type, it responds. No independent action. You’re in full control, and the AI just makes each step faster.
Examples: ChatGPT when you chat with it, GitHub Copilot suggesting code as you type, an AI email drafter, your AI assistant for business tasks. These tools are powerful, but they don’t do anything without you. They’re like a very fast intern sitting next to you, waiting for the next instruction.
Workflow agents: it follows a path you set
You design the steps. The agent runs them. If the process is “when a lead fills out a form, enrich the data, score it, and email the sales rep,” the agent handles all three steps. But you defined the path, and it doesn’t stray.
Examples: intelligent workflow automation tools like Zapier AI, Make, n8n with AI nodes, content pipelines, CRM automations. These are agentic workflows in practice. You build the track, the agent runs on it.
Autonomous agents: it decides the steps
You give a goal. The agent figures out how. It picks its own tools, decides the order, and loops until it’s done (or gives up). This is what most people picture when they hear “AI agent.”
Examples: Claude Code, Devin, research agents that browse the web and write reports, the kind of systems built on agentic AI frameworks. You trust it with a goal, and it goes.
Think of it like hiring. An assistant is an intern you supervise on every task. A workflow agent is a reliable process you’ve documented so well that anyone can follow it. An autonomous agent is a senior hire you trust with “figure it out.”
My take: most businesses should start in the middle. Workflow agents give you 80% of the benefit with 20% of the risk. Autonomous agents are exciting, but the failure rates are brutal. More on that next.
What each type looks like in practice
The classic five and the practical three aren’t competing. They overlap. Here’s roughly how they map:
| Autonomy level | Classic type(s) | Real tools |
|---|---|---|
| Assistants | Simple reflex, model-based | ChatGPT, Claude chat, Copilot, Jasper |
| Workflow agents | Goal-based | Zapier AI, Make, n8n, AI-powered CRMs |
| Autonomous agents | Utility-based, learning | Claude Code, Devin, custom LangChain/CrewAI agents |
For real AI agent examples across all three levels, I wrote a separate deep dive. And if you’re comparing specific products, the best AI agents roundup covers the top picks.
One more type worth knowing: multi-agent systems. That’s when you chain several agents together, each handling a piece of a bigger job. A research agent finds information, a writing agent drafts, a review agent checks quality. Sounds great in theory. In practice, this is where complexity goes through the roof, because every agent adds a new point of failure.
Where each type fails
In production, things go wrong a lot. And the data on autonomous agents is especially rough.
Assistants break when you expect them to remember or take initiative. They don’t. Every conversation starts fresh (unless the tool has memory bolted on). And they only do what you ask, so if you forget a step, it forgets too.
Workflow agents are brittle. They work great until the world changes. An API updates, a data format shifts, and the whole workflow stops cold. They need ongoing maintenance, like any other system.
Autonomous agents break the most, because reliability drops as autonomy rises:
- Agents fail 70 to 95% of the time in production, according to a synthesis of CMU and MIT research by Fiddler AI. That’s not a typo.
- The math gets worse when you chain agents. Three agents at 70% reliability each: 0.7 x 0.7 x 0.7 = 34% end-to-end success.
- Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027.
- The “Agents of Chaos” paper (February 2026) documented 11 different failure modes in live lab experiments, including resource loops, false completion reports, and cascading failures when multiple agents interact.
Andrej Karpathy (OpenAI co-founder) put it bluntly: agents work when you can check their output. Code and math, where you run the test and see if it passes. They struggle with things like strategy and advice, where there’s no easy way to tell if the answer is right.
My take: the more rope you give an agent, the more ways it can hang itself. That’s not a reason to avoid agents. It’s a reason to start with less autonomy and earn your way up. If you want to understand how to build them well, the principles of building AI agents is a good starting point.
How to pick the right type for your task
Anthropic’s own research team says it plainly: for most use cases, a single well-prompted AI call does the job. You probably don’t need an agent at all.
Only 21% of enterprises have redesigned workflows for agents (McKinsey). Most companies aren’t ready for full autonomy. And only 15% of IT leaders are deploying fully autonomous agents (Gartner).
Three questions to ask before you pick:
1. Can I define the steps? If yes, use a workflow agent. If the path is too unpredictable, you might need an autonomous one. But most business processes can be defined, and a well-built workflow beats a flaky autonomous agent every time.
2. Can I check the output? If the result is easy to verify (did the email send? did the data land in the right field?), you can safely give more autonomy. If it’s hard to verify (is this strategy advice good?), keep a human in the loop.
3. What breaks if it gets it wrong? High stakes = less autonomy. An agent that messes up a blog draft wastes an hour. An agent that messes up your CRM data costs you deals. Match the risk to the rope.
If you’re wondering how to build AI agents once you’ve picked a type, or want to understand the difference between agentic AI vs generative AI, those go deeper on the building and the concepts. There’s also a growing AI agent marketplace if you’d rather buy than build.
The honest answer for most businesses: start with workflow agents. Get those running well. Only then look at giving your AI more freedom.
How I can help
The biggest mistake I see is picking the wrong level of autonomy. Founders jump straight to autonomous agents because the demos look amazing, then spend months debugging something that a simple workflow would have handled. The agentive AI concept is real, but matching it to the right job is where the value actually lives.
If you’re trying to figure out what kind of agent fits your situation, I help founders scope that out. Whether to build or buy, how much autonomy to give it. Happy to think it through together.
FAQ
What are the 5 types of agents in AI?
The five classic types are simple reflex agents (act on current input only), model-based reflex agents (track what they can’t see), goal-based agents (plan toward an objective), utility-based agents (pick the best outcome), and learning agents (improve over time).
These come from Russell and Norvig’s Artificial Intelligence: A Modern Approach, the standard AI textbook. They describe how agents process information, but they don’t tell you which one to build for a business problem. For that, sorting by autonomy level (assistant, workflow agent, autonomous agent) is more practical.
What are the 7 kinds of AI agents?
The seven-type lists add multi-agent systems and hierarchical agents to the classic five. Some lists include hybrid agents for eight. Multi-agent systems chain several agents together, each handling part of a bigger task. Hierarchical agents have a “manager” agent that delegates to “worker” agents underneath. These extensions are real, but they add complexity fast. The compounding failure math above explains why.
Is ChatGPT an agent or LLM?
ChatGPT is a large language model (basically a very good text predictor) with some agentic features added on top. It becomes more agent-like when it browses the web, runs code, or uses tools. On the autonomy scale, it’s an assistant: it helps, but it doesn’t take independent action unless you specifically set up a workflow for it. The line between “LLM” and “agent” isn’t binary anymore. It’s a spectrum, and ChatGPT sits closer to the assistant end.
What is an autonomous agent?
An autonomous agent is an AI system that takes a goal and figures out its own steps to get there, with minimal human input. You say “research these 50 companies and rank them by growth rate,” and it decides which tools to use, what data to pull, and how to present it. Examples: Claude Code (writes and debugs software), Devin (builds applications), and custom research agents. The trade-off: autonomous agents are the most powerful type but also the most unreliable. They fail 70 to 95% of the time in production.
Who are the Big 4 AI agents?
This usually refers to the four major companies building agent infrastructure: OpenAI (GPT, ChatGPT), Anthropic (Claude), Google DeepMind (Gemini), and Meta (Llama). They’re not “agents” themselves. They’re the companies building the models and platforms that agents run on. If you’re looking at the best AI agents to actually use, that’s a different question with a longer answer.