Quick answer
The classic taxonomy names five types of AI agents: simple reflex agents (act on "if this, then that" rules), model-based reflex agents (track the state of their environment), goal-based agents (choose actions that move them toward a goal), utility-based agents (pick the best option according to a scoring function) and learning agents (improve from experience) (Russell and Norvig, AIMA). When several agents split the work, it is a multi-agent system. Today's LLM-based agents are mostly goal-based agents with some utility-based reasoning.
What is an AI agent?
An AI agent is a program that perceives its environment and takes actions to achieve a goal. Russell and Norvig define an agent as anything that perceives its environment through sensors and acts on it through actuators (Russell and Norvig, AIMA). In software, the "sensor" is an email, a ticket, a database record or a search result, and the "actuator" is an API call, a message or a change in a system.
You deal with agents every day. A motion sensor that turns on the hallway light is the simplest kind. A spam filter that learns each time you mark a message as junk or rescue one from the spam folder is a learning agent.
The types differ in one thing: how the agent makes a decision. That translates directly into cost, predictability and which business processes a given type can handle. For more on agency itself and how the agent loop works, see what is agentic AI.
The five types of AI agents at a glance
| Type | How it decides | Memory | Business example | When to choose it |
|---|---|---|---|---|
| Simple reflex | An "if condition, then action" rule | None | A ticket with "invoice" in the subject goes to the accounting queue | Predictable conditions, cheap mistakes |
| Model-based | A rule plus knowledge of the current state | Environment state | SLA monitoring that knows how often a ticket was reopened and how long it has waited | History and context matter, but no planning is needed |
| Goal-based | Picks actions that move toward a goal | State and goal | A helpdesk agent tasked with getting a ticket closed | Steps depend on the situation |
| Utility-based | Compares options with a utility function | State, goal, criteria | Delivery route planning that trades off cost, time and deadlines | Several conflicting criteria need weighing |
| Learning | Improves decisions from feedback | Experience | Product recommendations, fraud scoring, spam filtering | Lots of data, changing environment |
The types are not mutually exclusive. A utility-based agent also has a model of the world, and a learning agent may be goal-based inside. The taxonomy is really about the main decision mechanism.
Simple reflex agents: rules without memory
A simple reflex agent follows an "if condition, then action" rule and remembers nothing that happened before. A motion-sensor light turns on when it detects movement and off when movement stops. It does not predict when you will be back and does not learn your habits.
In a business, these are mail routing rules, threshold alerts and out-of-office auto-replies. They are fast, cheap and fully predictable. The downside is rigidity: if the sensor keeps firing on tree branches outside the window, it will keep doing so until someone changes the rule.
This is not a "worse" type. In many processes a rule is enough and beats a language model, because it is cheaper and easy to verify. Classic process automation (RPA and BPA) relies heavily on such rules, as we describe in our article on RPA and BPA automation.
Model-based reflex agents: decisions with state
A model-based reflex agent keeps an internal picture of its environment and updates it from observations. A robot vacuum builds a map of your home: it remembers where it has been, where the couch is and which rooms are left, even when it cannot see them.
That lets it handle an environment it cannot fully observe. In a business, these are systems that connect the current event with its history: ticket monitoring knows this customer is writing for the third time about the same issue and escalates differently than on first contact. It still reacts moment to moment, without planning several steps ahead.
Goal-based agents: acting toward a goal
A goal-based agent evaluates possible actions by whether they bring it closer to its goal and can plan a sequence of steps. A fitness app given a target weight builds a training and nutrition plan, and when you skip a workout, it reshuffles the following days.
This is the type closest to today's LLM-based agents. The agent gets a goal ("close this ticket," "prepare a report from these sources"), breaks it into steps, picks tools and judges whether it is getting closer. Tools are increasingly connected through the open MCP standard.
The limitation of a classic goal-based agent: it cares whether the goal is reached, not necessarily how well. A convenient path and an inconvenient one look equally good if both get there.
Utility-based agents: choosing the best option
A utility-based agent does not just ask "will I reach the goal?" but "which option is best?" It uses a utility function, a way of scoring outcomes that weighs several criteria at once. A building energy management system balances comfort, electricity cost and emissions instead of just holding a set temperature.
In a business, this means delivery route optimization (fuel cost, time, delivery windows), production scheduling, ordering a ticket queue by urgency and customer value, or dynamic pricing in e-commerce, which IBM cites as a typical utility-based agent (IBM).
The hard part is designing the utility function. Weigh the criteria wrong and the agent will consistently optimize for the wrong thing. That is why the business, not the technical team alone, should set the weights.
Learning agents: improving from experience
A learning agent changes its behavior based on feedback. A music streaming service starts with generic suggestions, but with every play, skip and playlist it understands your taste better. A classic learning agent has a critic (evaluates outcomes), a learning element (improves the rules) and a problem generator (suggests new things to try) (Russell and Norvig, AIMA).
In a business, these are recommendation systems, fraud detection, demand forecasting and spam filters. Their strength is handling environments that cannot be programmed in advance. Their weakness: they need lots of data, perform worse at first, and data quality determines result quality.
An important clarification: an LLM-based agent in production is usually not a learning agent in this sense. The model does not update its weights after each task. The agent "remembers" through notes, history and context, and quality improvements come from the team reviewing traces, refining tools and instructions, and sometimes fine-tuning the model.
Multi-agent systems: when agents split the work
A multi-agent system is several agents, each responsible for part of a larger task. In a smart home, rules turn on the lights, a model-based agent controls heating based on who is home, a goal-based agent schedules appliances for off-peak rates and a learning agent predicts preferences.
With language models, a popular pattern is the orchestrator: a lead agent splits the task, launches helper agents in parallel and then combines their results. Anthropic described such a system for research: on its internal evaluation the multi-agent version outperformed a single agent by 90.2%, but it uses about 15 times more tokens than a regular chat (Anthropic, 2025). The same write-up notes that tasks requiring shared context or with many dependencies, including most coding tasks, are a poorer fit for this kind of split.
| Approach | Pros | Cons | Good for |
|---|---|---|---|
| Single agent | Simpler, cheaper, easier to debug | Limited context, sequential work | Most business processes |
| Multi-agent system | Parallel work, specialization, more total context | Many times the cost, harder coordination | Research across many sources, easily divisible tasks |
You can see this in our AI agents: the helpdesk, the operations assistant, the outreach agent and the pentester are separate, specialized roles built on a shared agent-core. Each has its own narrow tool list and its own rules, and a person approves sensitive actions. Specialization makes control easier: an agent that does one thing is easier to evaluate and secure.
How do you choose the right type of agent for a business process?
Start with the process, not the technology. A few questions sort it out:
- Are the steps always the same? Yes: rules or a workflow (a reflex agent). No: a goal-based agent.
- Does the history of a case matter? Yes: you need a model of the world, meaning state and memory.
- Do you need to weigh conflicting criteria? Yes: a utility-based agent, with the weights set by the business.
- Do you have plenty of historical data and a changing environment? Yes: consider a learning component.
- Can the task be split into independent parts? Yes, and the value justifies the cost: a multi-agent system. No: a single agent.
In practice, the best systems combine types: rules where they suffice, an LLM agent where flexibility is needed, and a person where an important decision is made. We also cover the difference between a single agent and the broader agentic approach in AI agents vs agentic AI.
