Types of AI agents

AI agents are categorized by their architecture and capability level: simple reflex agents react to current inputs, model-based agents maintain an internal state, goal-based agents plan toward objectives, utility-based agents optimize across competing goals, and learning agents update their behavior from experience. In practice, production systems combine elements of several types.

Architectural categories

The classical AI taxonomy describes five agent types by increasing capability. Simple reflex agents map inputs directly to actions without memory — a thermostat is the canonical example. Model-based reflex agents maintain an internal model of the world to handle situations where the current input alone is insufficient to decide. Goal-based agents search for action sequences that lead to a defined goal state. Utility-based agents select actions by estimating expected utility, allowing trade-offs between competing objectives. Learning agents improve their behavior through feedback over time. Most production AI agents are hybrid: they have a language model for reasoning (goal-directed), access to memory or retrieval (model-based), and are fine-tuned from feedback (learning).

Functional categories in current practice

A more practical taxonomy focuses on what agents do rather than their internal architecture. Task automation agents execute defined workflows — retrieving data, processing it, writing results to a destination. Conversational agents interact with users in natural language while optionally taking actions. Research agents gather and synthesize information from multiple sources. Coding agents read and write software. Orchestrator agents decompose complex goals into sub-tasks and delegate to specialist agents. These functional categories are not mutually exclusive, and most real deployments are purpose-built combinations rather than pure instances of any single type.

Autonomy level as a dimension

Beyond architecture and function, agents vary by autonomy level — how much they act without human review. Fully supervised agents present every planned action for approval before execution. Semi-autonomous agents act independently for defined low-stakes actions and escalate high-stakes ones. Fully autonomous agents execute without per-action human review, relying on pre-defined permission scopes and post-hoc audit. The appropriate autonomy level depends on the reversibility of actions, the cost of errors, and the reliability of the agent on the specific task. Moving from supervised to autonomous deployment requires demonstrated reliability at the lower autonomy level first.

Types of AI agents — FAQ

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

A single-agent system uses one reasoning loop with access to multiple tools. A multi-agent system coordinates multiple agents, each with their own scope and capabilities, typically managed by an orchestrator that delegates sub-tasks. Multi-agent systems handle more complex tasks and enable parallel execution but introduce coordination overhead and new failure modes at agent boundaries.

Are large language models a type of AI agent?

A language model is a component an agent uses for reasoning, not an agent itself. An agent is a system that perceives its environment, reasons about it, takes actions, and receives feedback. A language model in a chat interface responds to queries but does not take actions in the world or maintain state across a task. When a language model is given tools, memory, and a task-execution loop, the resulting system is an agent.