Types of AI Agents: A Bridge from Classical Taxonomy to LLM-Based Agents (2026)
Two taxonomies coexist in the literature and rarely meet on the same page. The classical AI taxonomy, anchored in Russell and Norvig, defines six agent architectures. The modern LLM-era taxonomy, emerging in vendor and arXiv literature since 2022, defines five working categories. This page covers both and maps between them.
The taxonomies are not competing schools. They describe the same idea at different levels of generality. The classical taxonomy is a typology of decision-making architectures. The modern taxonomy is a typology of production patterns built on top of large language models. Most modern agents inherit from the classical lineage; mapping each modern type back to its classical parent is the editorial gap on this SERP.
Stimulus to action via a fixed rule. No internal state.
One model call, text in and text out. No tools, no memory.
Model can call functions to read or write outside the prompt.
Iterates the loop, often with chain-of-thought or self-critique.
Decomposes first, then runs each step. Two distinct calls.
Several agents, usually under an orchestrator. Parallel work.
Classical taxonomy
The six classical agent types, in order of increasing capability. Each is a self-contained architecture with its own decision-making procedure.
Simple reflex agent
An agent that selects an action based only on the current percept, ignoring all history. Behaviour follows hand-coded condition-action rules. A thermostat is the canonical example.
Source: Russell and Norvig, 4th ed., Ch. 2.
Model-based reflex agent
Maintains internal state to represent parts of the world it cannot currently see. The state is updated by a model of how the world evolves and how its actions change it. A self-driving car deciding to brake before it sees the brake lights of the next car ahead.
Source: Russell and Norvig, 4th ed., Ch. 2.
Goal-based agent
Acts to achieve an explicit goal. Search and planning algorithms select actions whose effects bring the agent closer to the goal. A delivery-routing agent computing the shortest path.
Source: Russell and Norvig, 4th ed., Ch. 2.
Utility-based agent
Picks the action that maximises a utility function over outcomes. Used when goals conflict or when goals can be achieved with varying quality. Most economic decision agents are utility-based.
Source: Russell and Norvig, 4th ed., Ch. 2.
Learning agent
Improves its performance by learning from experience. The architecture has a learning element, a performance element, a critic, and a problem generator. Modern reinforcement-learning agents are the contemporary instance.
Source: Russell and Norvig, 4th ed., Ch. 2.
Hybrid agent
Combines multiple architectures, typically a fast reactive layer for tight control loops and a slow deliberative layer for planning. Robotics and autonomous systems use hybrid architectures by default.
Source: Pattern documented across the BDI tradition; see also Wang et al. 2024 survey for the modern reading.
Most modern agents inherit from the classical lineage; mapping each modern type back to its classical parent is the editorial gap on this SERP.
Modern LLM-based agent types
The five categories that describe what vendors actually ship in 2026. The boundaries blur in practice; production systems often combine two or three.
Pure-LLM agent
An agent built around a single language-model call with no tools and no memory beyond the context window. The simplest 2024-era chatbot with retrieval falls into this category.
Source: Schluntz, Building effective agents (Anthropic, 2024) calls this an augmented LLM at the lowest tier.
Adds function calling, retrieval, or MCP-mediated tool calls to a language model. The agent reads from and writes to systems outside the model. The dominant production architecture in 2026.
Source: Yao et al., ReAct (arXiv:2210.03629) is the canonical paper for the reasoning-and-acting interleaving pattern.
Multi-step agent
Iterates the loop multiple times before producing a final result, often with chain-of-thought prompting and a self-critique step. Reflexion (Shinn et al., 2023) and Self-Refine (Madaan et al., 2023) are the cited prototypes.
Source: Shinn et al. 2023, Reflexion (arXiv:2303.11366); Madaan et al. 2023, Self-Refine (arXiv:2303.17651).
Planner-executor agent
Decomposes the goal into sub-tasks, then executes the sub-tasks with tool calls. Often the planner is a separate model call from the executor. Used when the task structure is non-trivial.
Source: Wang et al. 2024 survey (arXiv:2308.11432) catalogues planner-executor variants.
Multiple specialised agents collaborating, usually under an orchestrator. Useful when work is genuinely parallel or specialisable. Often overkill when a single agent with better tools would suffice.
Source: See dedicated page on multi-agent systems.
The mapping table
Reading both columns together is the point of this page. Vendor literature treats each modern type as a novel architecture. The classical lineage is older than the LLM. The mapping shows where each modern pattern descends from.
| Classical type | Modern equivalent | Where the lineage shows |
|---|---|---|
| Simple reflex agent | One-shot LLM caller (no memory, no tools) | Direct stimulus-response. The percept is the prompt; the action is the completion. |
| Model-based reflex agent | Pure-LLM agent with system prompt and conversation history | The system prompt and recent turns serve as the internal state model. |
| Goal-based agent | Planner-executor agent | The planner defines the goal-decomposition tree; the executor moves the agent toward leaves. |
| Utility-based agent | Reflection / self-refine agent | The critic step scores candidate outputs against an implicit utility function. |
| Learning agent | Reinforcement-learning-from-human-feedback (RLHF) trained model in agent loop | The model itself is the learned performance element; production training pipelines are the learning element. |
| Hybrid agent | Multi-agent system with a fast router and slow deliberative workers | The orchestrator runs reactive routing while specialised workers run deliberative planning. |
Why the taxonomy matters for procurement
When a vendor sells you an "AI agent", the architectural type determines the failure modes you will see in production. A pure-LLM agent with retrieval is a chat surface that hallucinates; the failure is wrong answers delivered confidently. A planner-executor agent with tool use is a system that mis-decomposes goals; the failure is busy work, partial completion, or runaway cost. A multi-agent system is a coordination problem; the failure is agents talking past each other.
A customer support deflection bot is typically a pure-LLM agent with retrieval. An enterprise workflow agent is typically a planner-executor with tool integration. A research-assistant agent is typically a multi-step agent with reflection. Knowing the type tells you which evaluation criteria matter, which questions to ask the vendor, and which parts of the procurement contract need to specify limits. The full evaluation framework is on how to evaluate an AI agent.