The types of AI agents most commonly used in artificial intelligence are simple reflex, model-based reflex, goal-based, utility-based, and learning agents. These five classical types describe how an agent selects actions. Modern lists often add LLM agents, hierarchical agents, and multi-agent systems, but those are better understood as implementation or orchestration architectures that can contain one or more classical decision types.

This guide will explain what AI agents are and how they work, as well as the types of AI agents with examples (covering the classical type of AI agents and modern agent architectures that extend them).

TL;DR

The canonical five types of AI agents are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents, and learning agents.
A simple reflex agent follows current-input rules. A model-based reflex agent also maintains internal state about what it cannot currently observe.
A goal-based agent plans toward a success condition, while a utility-based agent ranks possible outcomes when cost, risk, speed, or quality must be traded off.
A learning agent improves a decision component from feedback, but it does not necessarily retrain continuously in production.
LLM agents, hierarchical agents, and multi-agent systems describe modern agent architecture and coordination patterns rather than mutually exclusive sixth, seventh, or eighth classical types.
Use the least complex architecture that reliably meets the task, evaluation, security, and human-in-the-loop governance requirements.

What is an AI agent?

An AI agent is a system that perceives its environment, selects actions to achieve a performance objective, and affects that environment through software tools or physical actuators.

In classical artificial intelligence, an intelligent agent receives percepts through sensors and acts through actuators. Its agent function is the abstract mapping from percept history to action, while the agent program implements that mapping on a particular AI agent architecture.

The PEAS framework is a way to specify an agent before choosing its implementation:

Performance measure: the outcome to optimize, such as safe delivery, resolved cases, low fraud loss, or high schedule adherence.
Environment: the physical or digital world in which the agent operates.
Actuators: the mechanisms used to act, such as APIs, database writes, robot motors, emails, or ticket updates.
Sensors: the inputs used to perceive state, such as cameras, events, user messages, telemetry, or tool responses.

Modern AI agents add planning and reasoning, memory and tool use, retrieval, evaluators, and guardrails. The phrase sensors and actuators still applies as software agents sense through messages, events, APIs, and tool results, then act through software interfaces.

How do AI agents work?

AI agents work as a closed decision loop. They perceive, update state, decide, act, observe the result, and evaluate what should happen next. Not every agent performs every stage. A simple reflex system can skip explicit memory and planning, while an LLM agent may use retrieval, tools, evaluators, and several planning iterations.

A practical six-stage loop is:

Perception: sensors or interfaces receive an event, message, observation, or tool result.
State and memory: the agent updates its internal state or world model with information relevant to the current task.
Decision: a rule, policy, plan, utility function, or learned model selects the next action.
Action: actuators or tools change the environment, such as moving a robot, calling an API, updating a record, or requesting approval.
Observation: the agent checks the environment rather than assuming that the action succeeded.
Evaluation and learning: a critic, validator, test, reward signal, or human reviewer assesses progress and may trigger correction or model updates.

The task environment determines the complexity of the loop. A rule-only agent can work well when inputs are complete, and outcomes are predictable. A partially observable process requires state. A stochastic process can require expected-utility reasoning. A changing process may need a controlled learning pipeline.

Production controls sit around the loop. They include least-privilege credentials, tool allowlists, schema validation, transaction limits, sandboxes, audit logs, stop conditions, and escalation rules. These controls prevent a capable reasoning model from turning one incorrect assumption into a sequence of high-impact actions.

How many types of AI agents are there?

There are five classical types of AI agents in the widely used agent-structure taxonomy. Types include simple reflex, model-based reflex, goal-based, utility-based, and learning agents.

But these categories are not always mutually exclusive, as a learning agent can also be goal-based. A multi-agent system can contain utility-based agents. An LLM agent can be organized as a hierarchical agent with a manager that delegates work to specialist subagents.

Common count	Classification axis	Usually includes	Why the count differs
Four	Decision behavior	Reflex, model-based, goal-based, utility-based	Learning is treated as a capability layered onto the other types
Five	Classical agent structure	Simple reflex, model-based reflex, goal-based, utility-based, learning	Canonical structure used in many AI textbooks
Seven or eight	Mixed taxonomy	Five classical types plus hierarchical, multi-agent, or LLM agents	Behavioral types are combined with system architectures

Types of AI Agents at a Glance

The following table compares different types of AI agents in artificial intelligence and three modern architectures. The first five rows classify decision behavior. The last three classify implementation and coordination.

Type	Decision basis	Memory and planning	Learns?	Best fit and example	Main limitation
Simple reflex agent	Current percept and condition-action rule	No persistent state; no planning	No	Stable rules; thermostat or threshold alert	Fails when context is hidden, or rules change
Model-based reflex agent	Current percept plus internal state	Maintains a world model; limited look-ahead	Not required	Partial observability; robot vacuum or session-aware fraud check	Stale or incorrect state causes bad actions
Goal-based agent	Whether an action sequence reaches a goal state	Search or planning over future states	Not required	Route planning or task scheduling	Does not rank equally successful plans well
Utility-based agent	Expected utility across possible outcomes	Plans under uncertainty and trade-offs	Not required	Dispatch, resource allocation, risk-aware navigation	Utility misspecification can optimize the wrong outcome
Learning agent	Feedback used to improve a decision component	Depends on the underlying agent	Yes	Recommendations, adaptive detection, control	Drift, unsafe exploration, or poisoned feedback
LLM agent	Language-model policy plus tools and feedback	Often uses memory, retrieval, and iterative planning	Optional	Support resolution or repository coding	Hallucinated actions and compounding tool errors
Hierarchical agent	Manager decomposes and delegates work	Plan and state distributed across levels	Optional	Complex work split among specialist subagents	Bad decomposition or routing propagates downstream
Multi-agent system	Multiple agents coordinate or compete	Shared, private, or synchronized state	Optional	Parallel research, negotiation, distributed control	Communication overhead and conflict resolution

What are the five classical types of AI agents?

The five classical types of agents in AI explain how an individual agent converts percepts into actions.They form a capability progression from immediate reaction to state tracking, planning, trade-off optimization, and improvement from feedback.

1. Simple reflex agent

A simple reflex agent chooses an action from the current percept using condition-action rules. It does not retain a history, estimate hidden state, or evaluate long-term consequences. The logic is usually explicit: if condition X is observed, execute action Y.

A thermostat is the standard real-world example. If the measured temperature falls below the configured threshold, the controller turns on heating; if it rises above the threshold, it turns heating off. Similar patterns appear in spam rules, threshold-based monitoring, access-control policies, and basic game bots.

Simple reflex agents are valuable when a task is predictable, fully observable, low-risk, and latency-sensitive. They are easy to test because each rule can be checked against known inputs. They also provide deterministic behavior, which simplifies auditability.

Some of their limitations are that the same percept can require different actions depending on history or hidden context. A basic support router may send every message containing “refund” to the billing queue even when the customer is reporting refund fraud. In a partially observable task environment, current-input rules cannot distinguish these cases.

Choose a simple reflex agent when the rules cover the environment and consequences are local. Avoid it when the agent must remember prior events, reason about delayed effects, or recover from novel exceptions.

2. Model-based reflex agent

A model based reflex agent combines the current percept with an internal state that estimates what is happening in the environment. It updates that state using a world model, information about how the environment changes and how the agent’s actions affect it.

For example, a robot vacuum’s sensors cannot observe every room and obstacle at once, so the controller tracks visited areas, estimated position, blocked paths, battery level, and docking location. The next action can still be selected through rules, but those rules operate on the maintained state rather than on one sensor reading.

Digital systems use the same pattern. A fraud-review service may combine the current payment with session history, device identity, previous authentication results, and recent account changes. The current transaction alone is incomplete, the state makes the situation interpretable.

A model based reflex agent is suitable for a partially observable environment that remains predictable enough to model. It reacts faster than a full planner because it does not need to search many future action sequences for every decision.

Its main risk is state drift. If observations are missing, delayed, duplicated, or incorrectly linked, the internal state can diverge from reality. Mitigations include state expiration, reconciliation against authoritative systems, confidence estimates, event idempotency, and explicit handling for unknown state.

3. Goal-based agent

A goal based agent selects actions by evaluating whether they move the system toward a defined goal state. Unlike a reflex agent, it can compare future action sequences through search or planning before acting.

A route planner is a practical example. The current location does not determine one fixed action. The system considers available roads, transitions between locations, constraints, and a destination goal. A scheduling agent similarly searches for a sequence of bookings that satisfies required participants, duration, availability, and deadline constraints.

Goal-based behavior is appropriate when the success condition is clear, and the action path is not fixed in advance. It supports multi-step work such as research planning, configuration changes, order fulfillment, or software tasks where each action changes what should happen next.

The limitation is that goal completion is binary. Several plans may reach the same goal but differ greatly in safety, cost, latency, resource use, or user preference. A route that arrives at the destination technically succeeds even if it is expensive, risky, or much slower than an alternative.

Choose a goal based agents when the outcome can be expressed as a clear goal test, and the search space is manageable. When successful outcomes must be ranked across competing criteria, add a utility model rather than relying on goal completion alone.

4. Utility-based agent

A utility based agent assigns a numerical preference to possible outcomes and chooses the action with the highest expected utility. This extends goal-based reasoning by distinguishing among outcomes that all satisfy the goal.

Suppose an autonomous delivery system must choose among three routes. Every route reaches the destination, but they differ in travel time, collision risk, energy consumption, road restrictions, and delivery-window reliability. A utility function combines these criteria, while expected utility accounts for uncertain events such as traffic or weather.

Utility-based agents fit stochastic, multi-objective environments. Real applications include fleet dispatch, dynamic resource allocation, risk-aware pricing support, portfolio decision support, production scheduling, and robot navigation.

Here is a table that helps you decide when to choose a goal-based agent and a utility-based agent:

Dimension	Goal-based agent	Utility-based agent
Decision test	Does the plan reach the goal?	Which outcome has the highest expected value?
Treatment of trade-offs	Limited or encoded as constraints	Explicitly ranks cost, risk, speed, quality, or preference
Uncertainty	Can plan, but may not quantify outcome preference	Uses probabilities and expected utility when available
Primary risk	Chooses an inefficient but valid plan	Optimizes a poorly designed or incomplete utility function

The primary failure mode is utility misspecification. If the objective rewards speed while underweighting safety, the agent can rationally optimize for the wrong behavior. Controls include multi-objective constraints, hard safety limits, stakeholder review, offline simulation, sensitivity analysis, and monitoring for reward hacking or proxy gaming.

5. Learning agent

A learning agent improves its behavior from experience or feedback. Learning is an overlay rather than a completely separate decision mechanism; the performance element being improved may itself be reflex-based, model-based, goal-based, or utility-based.

The classical learning-agent architecture contains four components:^]

Performance element: selects external actions.
Learning element: modifies the performance element to improve future behavior.
Critic: evaluates results against the performance standard.
Problem generator: proposes informative actions or exploration that may improve learning.

For example, fraud models retrained on confirmed cases, adaptive traffic control, predictive maintenance, and reinforcement-learning controllers. However, a deployed learning agent does not have to update itself continuously. Many production systems learn offline as data is collected, validated, used in a controlled training pipeline, evaluated, approved, and then released as a new version.

Learning agents are appropriate when the environment changes, feedback is measurable, and experimentation can be made safe. Their risks include model drift, feedback loops, poisoned labels, unsafe exploration, catastrophic forgetting, and the optimization of short-term feedback at the expense of long-term outcomes. Versioned training data, holdout evaluation, drift monitoring, approval gates, and rollback are essential controls.

What modern agent architectures extend the five classical types?

Modern AI agent architectures extend the five classical types by adding language-model reasoning, specialized delegation, or coordination among multiple agents.

LLM and Tool-Using Agents

An LLM agent uses a large language model as a flexible policy or planner and connects it to memory, retrieval, and tools. The model interprets instructions, selects a tool, observes the result, and decides whether to continue, correct, escalate, or stop.

A production LLM agent typically includes:

A model and instruction policy.
Short-term task state and, where justified, long-term memory.
Retrieval from approved knowledge sources.
Typed tools with narrow permissions.
Planning or iterative action selection.
Validators, tests, or evaluator steps.
Stop conditions and human approval gates.

A real-world example is a coding agent that receives a repository issue, inspects files, edits code, runs tests, and opens a pull request for review. For example, GitHub documents agents can independently execute development tasks and return pull requests, while preserving human review as the acceptance boundary.

LLM agents fit open-ended language and tool workflows where the required steps cannot be hardcoded. Their risks are hallucinated tool calls, prompt injection, incorrect assumptions, high latency, and compounding errors. Tool schemas, sandboxing, environmental feedback, and verification are more important than adding more reasoning tokens.

Hierarchical Agents

A hierarchical agent organizes work across levels. A manager or orchestrator decomposes a goal, delegates subtasks to specialist subagents, tracks progress, and aggregates results. The workers can use different models, tools, knowledge bases, permissions, or classical decision types.

For example, Microsoft’s orchestrator-subagent architecture uses a primary agent to delegate work to specialists and recommends it when a problem has clear functional boundaries, independent ownership, or reusable domain expertise.

This architecture fits tasks that are too broad for one prompt or toolset but can be decomposed. The main risk includes a poor task breakdown or routing decision that sends downstream agents the wrong problem. Use typed task contracts, bounded recursion, progress checks, shared identifiers, and manager-level validation.

Multi-Agent Systems

A multi-agent system contains multiple autonomous or semi-autonomous agents that cooperate, negotiate, compete, or review one another. Coordination can be sequential, concurrent, conversational, or based on handoffs.

A multi-agent research workflow might assign separate agents to literature search, source verification, quantitative analysis, and synthesis. A concurrent security workflow might ask independent agents to review code for different vulnerability classes, then aggregate findings. These designs are useful when specialization or parallel work provides measurable value.

The risks include duplicated effort, contradictory outputs, communication loops, shared-state conflicts, and unclear accountability. Effective agent orchestration requires explicit turn-taking or routing, conflict-resolution logic, state ownership, time and cost budgets, and one component responsible for final acceptance.

Most production systems are hybrid. They combine deterministic rules for safety-critical boundaries, model-based state, goal or utility planning, learned models, LLM reasoning, specialist agents, and human approval. For a deeper implementation discussion, see our AI agent routing guide.

Which Type of AI Agent Should You Choose?

Choose the least complex type that can reliably handle the observability, planning horizon, uncertainty, adaptation, and coordination requirements of the task.

Use this decision sequence:

Are the rules complete and is the current input sufficient? Use a simple reflex agent.
Does the correct action depend on hidden or historical state? Use a model-based reflex agent.
Must the system select a multi-step path to a clear success condition? Use a goal-based agent.
Do several acceptable outcomes need to be ranked by cost, risk, time, or quality? Use a utility-based agent.
Must behavior improve because patterns change or the designer cannot specify the policy fully? Add a learning agent capability.
Can the work be divided into specialist domains under one manager? Use a hierarchical agent.
Does the task benefit from independent parallel analysis, negotiation, or distributed control? Use a multi-agent system.
Is the workflow open-ended, language-heavy, tool-enabled, and difficult to hardcode? Use an LLM agent or hybrid architecture.

Types of AI Agents With Examples: Real-World Applications

To understand types of AI agents with examples, it is best to examine how each architecture observes its environment, makes decisions, takes action, and evaluates the result. Production systems frequently combine several AI agent types, so the classifications below identify the dominant behavior rather than treating each application as a pure implementation of one category.

Rule-based environmental control (simple reflex agent): A simple reflex agent reacts to the current sensor reading using predefined condition–action rules. For example, a conventional thermostat activates heating when the measured temperature falls below a threshold and switches it off when the target temperature is reached.
Payment-fraud screening (model-based reflex agent): A model-based reflex agent evaluates the current transaction together with information about the customer, device, merchant, account history, and related transactions. This stored world model allows the system to distinguish between superficially similar payments occurring under different circumstances.
Last-mile delivery routing(goal-based or utility-based agent): A goal-based agent searches for actions that lead to a defined outcome, such as completing every delivery. A more advanced utility-based agent also compares valid routes according to distance, fuel consumption, delivery commitments, and operational cost.
- For example, UPS uses its ORION routing system to determine the order in which drivers make deliveries and pickups. UPSNav then provides detailed directions based on those optimized routes. UPS reports that the system is designed to reduce miles driven, fuel consumption, and emissions while improving delivery service.
Data-center cooling(utility-based learning agent): A data-center cooling controller can function as both a learning agent and a utility-based agent. It learns how changes to fans, cooling equipment, and airflow affect future temperatures, then selects actions intended to maintain safe conditions while minimizing energy use.
- Google researchers deployed a model-based reinforcement-learning system to regulate temperature and airflow in a large-scale data center. Learning from operational data, the system safely improved efficiency over existing controllers.
Personalized content recommendation (learning agent): A recommendation engine is a learning agent because it improves its predictions using behavioral feedback. It observes interactions such as viewing history, search activity and engagement, ranks available content, and updates its models through controlled offline training and online experimentation.
- Netflix’s recommendation system uses multiple algorithms to personalize the member experience. Its published research explains that recommendation changes are evaluated using historical engagement data and controlled A/B testing focused on outcomes such as retention and medium-term engagement.
Software issue resolution (LLM agent): An LLM agent can inspect a software repository, navigate files, edit code, execute commands and run tests while working toward the goal described in an issue report.
- For example, the SWE-agent research project converts language models into software-engineering agents that attempt to resolve issues in real GitHub repositories. Its agent architecture provides specialized commands and feedback interfaces for browsing, editing and executing code. The associated work was presented at NeurIPS 2024.
Legal contract review (hierarchical agent): A hierarchical agent divides a large review into smaller specialist tasks. A coordinating agent might assign clause extraction, policy comparison, risk classification and summary generation to separate components before assembling the results for a lawyer.
Collaborative security analysis (multi-agent system): A multi-agent system assigns different parts of an investigation to specialized agents. One agent may inspect network activity, another may evaluate source code, and another may correlate identity or endpoint evidence. An orchestrator then reconciles their findings.
Customer-support resolution (hybrid LLM agent): A hybrid support system combines an LLM agent, retrieval, deterministic business rules, API tools and human-in-the-loop governance. The language model interprets the request and plans the workflow, while policy controls determine which actions it may execute and when escalation is mandatory.
- For example, Intercom’s Fin agent resolves customer requests across support channels and can automate workflows such as refunds, account changes and subscription updates. Intercom also retains human escalation paths for cases that require additional judgment or authority.

Want more real-world agentic AI examples?

100+ Agentic AI Use Cases

Conclusion

The five classical types of AI agents define individual decision-making: reacting, tracking state, planning, optimizing utility, or learning. Modern LLM agents, hierarchical agents, and multi-agent systems explain implementation and coordination.

Choose the least complex design satisfying the task environment and governance. Only add state for incomplete data, planning for sequential steps, utility for ranking outcomes, learning for controlled feedback, and multiple agents for measurable coordination value.

Frequently Asked Questions

What are the five types of AI agents?

The five classical types of AI agents are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents, and learning agents. They differ in whether decisions use only the current percept, internal state, future goals, ranked utility, or feedback that improves behavior.

Why do some sources describe seven or eight types of AI agents?

Seven- or eight-type lists usually mix classification axes. They add LLM agents, hierarchical agents, or multi-agent systems to the five classical decision types. These additions describe implementation and coordination, so they can overlap with reflex, goal-based, utility-based, or learning behavior.

What are the 4 types of agents in AI?

Four-type lists usually include reflex, model-based, goal-based, and utility-based agents while treating learning as a capability that can improve any of them. Other lists combine simple and model-based reflex agents. Always check which classification axis a source uses.

What is the simplest type of AI agent?

A simple reflex agent is the simplest type. It maps the current percept directly to an action through condition-action rules. It is fast and predictable but cannot use history, infer hidden state, plan ahead, or adapt when rules no longer match the environment.

What is the difference between a simple reflex agent and a model-based reflex agent?

A simple reflex agent acts only on the current percept. A model-based reflex agent maintains internal state and a world model, allowing it to act when the environment is partially observable. The additional state improves context but introduces synchronization and drift risks.

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

A goal-based agent checks whether a plan reaches a desired state. A utility-based agent also ranks successful outcomes according to cost, risk, time, quality, or other preferences. Utility is needed when several valid plans are not equally desirable.

Which type of AI agent can learn from experience?

A learning agent improves a performance element using feedback from a critic and, where appropriate, exploratory input from a problem generator. Learning may occur online or through a controlled offline training pipeline. The underlying decision system can still be reflex, goal-based, or utility-based.

Is an LLM agent a separate type of AI agent?

An LLM agent is better treated as a modern implementation pattern, not a universally accepted sixth classical type. The language model supplies flexible reasoning or planning, while memory, retrieval, tools, evaluators, permissions, and orchestration determine how the complete system behaves.

What is the difference between a hierarchical agent and a multi-agent system?

A hierarchical agent uses explicit levels, typically a manager delegating to subagents. A multi-agent system is broader: agents may cooperate, compete, work concurrently, debate, or hand off tasks without a strict hierarchy. A hierarchical design is therefore one form of multi-agent architecture.

Which AI agent type is best for customer service?

Customer service usually needs a hybrid LLM agent: language understanding for the conversation, model-based state for customer context, goal-based planning for resolution, deterministic policy rules, tools for approved actions, and human escalation for exceptions or high-impact decisions.

Are AI agents the same as agentic AI?

An AI agent is an individual system that perceives, decides, and acts. Agentic AI refers more broadly to systems or workflows that use one or more agents to complete goals with some autonomy. Agentic AI can include deterministic workflows, LLM agents, hierarchical delegation, and multi-agent coordination.

Types of AI Agents: How Each Type Works, Examples, and When to Use It