What Are the 7 Types of AI Agents? Plain English Guide
The seven types are simple reflex, model-based reflex, goal-based, utility-based, learning, hierarchical, and multi-agent systems. The first five come from Russell and Norvig's Artificial Intelligence: A Modern Approach, the standard textbook on the subject. The last two come from industry practice rather than that textbook, and describe how agents are actually assembled in production today.
That list is worth knowing, and it is also mostly useless for a buying decision. Here is the list, then the part that matters.
The seven, in order of complexity
# | Type | What it does | Where you see it |
|---|---|---|---|
1 | Simple reflex | Follows if this, then that. No memory | Invoice routing, form triage, guards |
2 | Model-based reflex | Keeps a picture of current state, not just the input | Assistants that track an open case |
3 | Goal-based | Plans steps to reach a stated goal | Tool-using agents, multi step tasks |
4 | Utility-based | Optimises a trade off, not just success or failure | Cost and latency routing, scheduling |
5 | Learning | Improves from feedback over time | Exception handling that gets better |
6 | Hierarchical | A manager agent breaks work up and delegates | Where work is large enough to split |
7 | Multi-agent | Several agents cooperate or compete | Cross functional processes |
Why the list is less useful than it looks
Every agent you would actually deploy is a blend. A working accounts payable agent is a simple reflex layer for the obvious matches, a goal-based layer for the messy invoices, a utility layer deciding when to spend more effort, and a learning layer that tightens the rules as your team approves or corrects output. Nobody buys one type.
So the honest answer to which type you need is: the smallest combination that covers your work, built so each part can be tested on its own.
The distinction that actually decides your outcome
The word agent is used for two different things, and the difference is worth money.
What to compare | Assistant | Agent |
|---|---|---|
Waits for you | Yes | No. It reads work as it arrives |
Has tools | Sometimes | Yes, and it uses them |
Remembers state | Within a conversation | Across the whole process |
Acts on its own | No, it suggests | Yes, within limits you set |
Runs when nobody is watching | No | Yes |
An assistant that suggests the right answer still leaves the work with your team. An agent that owns the role takes the work off them. Only the second one shows up as recovered hours, which is why the two get priced so differently.
The useful test is not which of the seven types a product claims. It is whether the thing reads every line of the work, or waits for someone to paste something into it.
Where this goes next
What does an AI agent do exactly?
How to decide which roles your AI agents should own
Should you hire another person or deploy an AI agent?
Frequently asked questions
How many types of AI agents are there? Seven in the common business taxonomy. Five are academic, taken from Russell and Norvig's Artificial Intelligence: A Modern Approach (simple reflex, model-based reflex, goal-based, utility-based, learning). Two are architectural: hierarchical and multi-agent systems.
What are the 5 types of AI agents? The classic five come from Russell and Norvig: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. The seven type list adds hierarchical agents and multi-agent systems.
Are hierarchical and multi-agent the same? No. A hierarchical setup has a manager agent that splits work and hands pieces to specialist agents. Multi-agent means several agents working together, and they can be peers rather than a hierarchy. Both add coordination problems, which is why small deployments rarely need either.
Which type should a business use? Almost always a combination of simple reflex for the clear cases, goal-based for the rest, and a learning layer on top. Start with the least autonomy that covers the work.
Do I need a multi-agent system? Only when the work genuinely splits into parts that need different skills. For one role on one workflow, a single agent plus a person approving exceptions is faster to build and easier to trust.
Is agentic AI the same as an AI agent? No. Agentic AI describes systems with agent-like behaviour. An AI agent is the specific thing doing the work. The terms get used interchangeably in marketing, which is usually a sign that nobody has defined the scope.
