Types of AI Agents: 10 Kinds Explained, With Business Examples
The five classic types of AI agent from AI textbooks, the five kinds businesses build with language models today, and how to tell which one a task needs.
What types of AI agent are there?
There are two useful ways to classify AI agents. The classic textbook types describe how an agent decides: simple reflex, model-based reflex, goal-based, utility-based and learning agents. The types businesses build today describe what a language-model agent does: conversational assistants, workflow agents, research agents, coding agents and computer-use agents. Most real systems combine traits from both lists, and many tasks sold as needing an agent are better served by a fixed workflow.
This explainer is for business and product leaders who keep hearing "agent" used for very different things. An AI agent, in the textbook sense, is anything that perceives its environment and acts on it. In current products, it usually means a language model that can use tools and take several steps towards a goal. For the difference from a chatbot, see AI agent vs AI chatbot.
The five classic types
These come from Stuart Russell and Peter Norvig's textbook Artificial Intelligence: A Modern Approach (chapter 2). They are still the clearest way to think about how much an agent knows and plans.
1. Simple reflex agents act only on what they perceive right now, using condition-action rules: if this, then that. A thermostat is the textbook example. Business example: a rule that routes any email containing "invoice" to finance. Fast and predictable, but helpless when the rule does not fit.
2. Model-based reflex agents keep an internal picture of the parts of the world they cannot currently see, updated as things happen. Business example: a support bot that remembers what the customer said earlier in the conversation and the state of their order, and responds accordingly.
3. Goal-based agents consider what their actions will lead to and choose actions that reach a goal, which means planning. Business example: an agent given "book a meeting with this client next week" that checks calendars, proposes times and sends the invite.
4. Utility-based agents go further and weigh how good each outcome is, trading off goals such as speed, cost and risk. Business example: a delivery-routing system that balances fuel cost, delivery windows and driver hours rather than finding any valid route.
5. Learning agents improve their behaviour from feedback over time. Business example: a recommendation system that updates what it suggests based on what customers buy. Language-model agents learn in a looser sense, through updated instructions, examples and evaluation, rather than retraining after every task.
The five kinds businesses build today
Modern agents are built around a language model that reads the situation, decides on a step, calls a tool and repeats. They are grouped here by the job they do.
6. Conversational assistants and support agents talk to customers or staff and can take limited actions, such as looking up an order or creating a ticket. Voice agents are the spoken version. Watch for: clear handover to people and limits on what they may promise. See AI voice agents for customer service.
7. Workflow agents carry out a multi-step business process, such as processing an inbound request through checks, systems and approvals, choosing the next step when cases vary. Watch for: approvals before irreversible actions. See agentic workflow automation.
8. Research and knowledge agents search documents, databases or the web, compare sources and write a summary with references. Watch for: checking citations, since models can misstate what a source says.
9. Coding agents read a codebase, make changes, run tests and propose pull requests. Watch for: sandboxing, permissions and code review. See securing AI coding agents.
10. Computer-use and browser agents operate software through its screen, clicking and typing like a person, which helps with systems that have no API. Watch for: slower, less reliable runs and strong limits on what they can access. See computer-use agents.
Worth noting
Multi-agent systems are not a separate type so much as an architecture: several of the agents above, each with a narrow role, coordinated by an orchestrator. See AI agent orchestration.
How the two lists connect
The classic types are a useful test of what a modern agent actually does. Many products marketed as agents are reflex agents with a language model inside: they react to one input and stop. That is often exactly what you want, because it is predictable. Goal-based behaviour, where the system plans several steps, is where both the value and the risk increase.
| Type | Decides by | Typical business use | Risk level |
|---|---|---|---|
| Simple reflex | Fixed rules on current input | Routing, alerts, validation | Low |
| Model-based reflex | Rules plus remembered state | Context-aware chat and support | Low to medium |
| Goal-based | Planning steps to reach a goal | Scheduling, multi-step tasks | Medium |
| Utility-based | Weighing outcomes against each other | Routing, pricing, allocation | Medium |
| Learning | Improving from feedback | Recommendations, ranking | Medium (drift) |
| Conversational/support | Model plus limited tools | Customer and staff questions | Medium |
| Workflow | Model choosing process steps | Request handling, back office | Medium to high |
| Research/knowledge | Model searching and summarising | Reports, due diligence | Medium |
| Coding | Model editing and testing code | Software changes | High without review |
| Computer-use | Model operating a screen | Legacy systems without APIs | High |
Which type does your task need?
- If the same input should always produce the same action, use rules (a simple reflex agent), not a language model.
- If the task needs reading unstructured text but the steps are fixed, use a workflow with one AI step.
- If the steps genuinely vary by case, consider a goal-based workflow agent with approvals and logs.
- If outcomes must be optimised across trade-offs, you need explicit scoring, not just a model's judgement.
- If no API exists, a computer-use agent may bridge the gap, but plan to replace it with an integration.
Pro tip
Our decision framework tests whether a process suits an agent at all, and AI agent guardrails covers how to limit what one can do.
Conclusion
Knowing the type of agent clarifies what you are buying or building: how much it plans, how much it can do on its own and how much can go wrong. Use the classic types to judge decision-making, and the modern types to describe the job. Start with the simplest type that does the task. For how to build one, see the AI agent development guide; for whether to build or buy, build vs buy AI agents.
Working out which kind of agent you need?
ZSpace Labs designs and builds AI workflows and agents connected to business systems, with permissions, approvals and monitoring. See our AI automation service.
Sources
Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th edition (Pearson, 2020), chapter 2, Intelligent Agents.
Common questions.
In Russell and Norvig's Artificial Intelligence: A Modern Approach, the classic types are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents. They describe how an agent decides, not what business task it does.