An AI agent is an AI system that uses observations or other context to choose and carry out actions toward an objective, with some degree of autonomy.
This is an operational definition for examining system behavior. The OECD's 2026 analysis of AI agents identifies common foundations while documenting differences in how definitions treat autonomy, outputs, and environments. The term does not specify one architecture.
An agent and a model are different objects
A model produces outputs from inputs, such as a score, generated text, or a proposed action. An agent connects a decision process to a way of acting in its environment. In a model-based agent, model inference contributes to that process; the surrounding execution mechanisms matter too.
For example, a model output that describes a file change does not itself establish that the file changed. The system needs a mechanism that applies the change. Confusing these objects makes it easy to review the quality of a suggestion while overlooking what can execute it.
How context becomes action
Observations give the agent information relevant to its objective. They may be incomplete or outdated. The decision process selects an action using the available context, and the action can produce new observations that affect a subsequent choice.
An agent does not have to use a large language model. Nor does every implementation express actions as external tool calls or require a particular protocol, such as MCP. A tool call is one way to connect a decision to an operation, not the definition of agency.
Autonomy has a scope
Hypothetical example: a navigation agent in a simulated arena receives its position and an obstacle map, selects a route toward a destination, and issues movement commands. It can choose another route when it encounters an obstruction. Its goal remains fixed, and the permitted movement stays within the arena. Choosing steps does not require authority to redefine the task.
To describe that autonomy precisely, state which choices the system makes, which require a person, and what ends its activity. A long runtime or a large number of actions cannot answer those questions alone. The label also says nothing about whether the selected actions are authorized, correct, or successfully completed.
Related concepts
Agentic AI describes how a workflow delegates choices and actions to agents.
AI agent governance defines the policies, decisions, controls, and evidence that govern an agent's delegated authority. It addresses what the agent may do, alongside the technical account of what it can do.
The AI agent governance pillar explains how to examine tools, execution paths, approvals, and evidence in an enterprise workflow.