Types of AI Agents: Five Classical Types With Examples

The primary types of AI agents comprise five classical architectures: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. Wikipedia lists this classification under Russell and Norvig. The definitions and examples below come from IBM Think, Wikipedia and Anthropic Engineering.

According to IBM Think and Wikipedia, an artificial intelligence (AI) agent perceives its surroundings through sensors and acts through actuators to achieve specific outcomes. In multi-agent configurations, all five types can be deployed together in a multi-agent system, each handling the part of the task it suits.

Anthropic Engineering draws a separate distinction between workflows, where predefined code paths orchestrate LLMs and tools, and agents, where LLMs direct their own tool use in a loop.

The Five Types of AI Agents

The five classical types of artificial intelligence (AI) agents are simple reflex, model-based reflex, goal-based, utility-based, and learning agents. These architectures are ordered by how much environmental information they remember and evaluate before taking an action through their actuators.

Wikipedia counts anything from a basic thermostat to a human being as an intelligent agent. IBM Think sorts agents by level of intelligence, decision-making process and how they interact with their surroundings.

• Simple reflex agents: Execute predefined condition-action rules based solely on current sensor inputs without storing historical data.
• Model-based reflex agents: Maintain an internal representation of the world to track current conditions and past interactions in partially observable environments.
• Goal-based agents: Evaluate possible future states against explicit targets using planning and reasoning, often relying on decision trees.
• Utility-based agents: Score multiple potential outcomes against a utility function to resolve trade-offs among competing operational goals.
• Learning agents: Adapt and improve decision performance over time through dedicated feedback loops, critics, and exploratory action generators.

Simple Reflex and Model-Based Reflex Mechanisms

Simple reflex agents operate on immediate sensory inputs through fixed condition-action rules without evaluating past states or future consequences. According to IBM Think, a simple reflex agent acts strictly on current perceptions received through its sensors. A common household thermostat demonstrates this mechanism by activating heat when ambient temperature falls below a set threshold and deactivating power when the target is reached.

An automatic traffic light provides another example by adjusting to traffic sensor inputs without reference to past states. Because simple reflex systems store no historical information, they perform effectively only in structured, predictable environments with well-defined rules. They can repeat mistakes.

Model-based reflex agents still use condition-action rules but also keep an internal model of the world. That model tracks the current state of the environment and accounts for how past interactions altered surroundings. For instance, a robot navigating a room uses its internal model to consider previous movements and track the locations of obstacles already met.

IBM Think notes that an internal model allows agents to function in partially observable environments where sensors cannot capture the full operational picture. These systems still rely on condition-action rules, which leaves them unable to perform advanced reasoning or adaptive learning in dynamic environments. For a comparison of deterministic scripts and autonomous systems, see AI automation vs AI agents.

Goal-Based and Utility-Based Decision Frameworks

Goal-based agents select actions by evaluating future states against explicit objectives, whereas utility-based agents weigh trade-offs among multiple competing outcomes using a utility function. A goal-based agent moves beyond reflex triggers by establishing an explicit goal, examining possible actions, and selecting the path most likely to reach that target. According to IBM Think, these agents rely on planning and reasoning to assess future consequences.

A robot navigating a building illustrates the goal-based approach by planning a path to a specific room while avoiding known obstacles and minimizing detours. Because many goal-based systems rely on preprogrammed strategies or decision trees, their operational flexibility remains limited when environments scale in complexity. Applications for goal-based architectures include robotics, autonomous vehicles, and complex simulation platforms.

IBM says a utility-based agent assigns a value to each possible outcome and picks the action with the highest overall utility, which suits tradeoffs between multiple goals. Building accurate, reliable utility functions is challenging.

IBM Think uses a self-driving car as a utility-based example, where the system must balance speed, fuel efficiency and safety. Similarly, an e-commerce utility-based agent sets prices and recommends products from sales history, customer preferences and inventory levels. Engineering teams evaluating autonomous decision layers can see autonomous AI agents for business.

The Four Components of Learning Agents

A learning agent improves its operational performance over time by acquiring knowledge from feedback and environmental interactions rather than relying entirely on static rules. IBM Think outlines that learning agents operate using four structural components.

IBM describes reinforcement learning as an agent that tries strategies, receives rewards and penalties, and refines its approach. Applications cited by IBM include autonomous driving systems, robotics, and virtual assistants that help human agents in customer support.

The four components are:

• Performance element: Chooses actions from a knowledge base.
• Learning element: Adjusts knowledge from feedback and experience.
• Critic: Evaluates action quality, often as rewards or penalties.
• Problem generator: Suggests exploratory actions.

Multi-Agent Systems and the Smart-Factory Example

In IBM's description, hierarchical agents break complex problems into subtasks. Higher-level agents focus on overarching goals and lower-level agents handle specific tasks. All five types can be deployed together.

A smart factory is IBM's example of all five working together:

• Simple reflex agents: Handle basic automation such as shutting down a conveyor belt on a safety hazard.
• Model-based reflex agents: Track machine states and recognize maintenance needs before failure.
• Goal-based agents: Drive goals such as optimizing production schedules or reducing waste.
• Utility-based agents: Weigh operational factors including energy consumption, cost efficiency, and production speed.
• Learning agents: Improve operations through reinforcement learning and machine learning.

Language Model Workflows and Dynamic Control Loops

Anthropic separates workflows from agents by who controls the flow: hardcoded software or the LLM. In technical analysis published on December 19, 2024, Anthropic Engineering groups both patterns together as agentic systems. Anthropic defines workflows as systems where LLMs and tools are orchestrated through predefined code paths.

In agents, models dynamically direct their own processes and tool usage and keep control over how they finish tasks. Anthropic observes that agents are typically just LLMs using tools based on environmental feedback in an iterative loop. This architecture suits open-ended tasks where the required sequence of steps cannot be predicted in advance and fixed pathways cannot be hardcoded.

This control-flow division is a different axis from the classical five types, which Wikipedia lists under Russell and Norvig and IBM Think describes. The classical taxonomy categorizes agents by internal decision models, whereas Anthropic categorizes systems by how control is governed. Sourced documentation does not map specific commercial language model products onto Russell and Norvig's five classical types.

Greater autonomy introduces higher operational expenses and increases the risk of compounding errors across multi-step execution paths. Anthropic recommends extensive sandbox testing and appropriate guardrails. Readers evaluating small-scale business deployments can see AI agents for small business.

Selection Criteria Across Types of AI Agents

IBM says simple reflex agents suit structured, predictable environments, model-based reflex agents suit partially observable ones, and utility-based agents suit dynamic, complex ones. Guidance from Anthropic Engineering recommends finding the simplest solution possible and adding architectural complexity only when necessary. For many operational use cases, optimizing single LLM calls with retrieval and targeted examples is usually enough without building multi-step agentic systems.

Agents are a poor fit when the steps can be written down in advance. Anthropic says that for many applications, optimizing a single LLM call with retrieval and in-context examples is usually enough, and that agents cost more and can compound errors. Anthropic's post is dated December 19, 2024 and says much of the tooling it describes has changed since.

Anthropic recommends extensive sandbox testing and appropriate guardrails for agents. Developers structuring custom implementations can see AI agent development for small business.