Autonomy

Autonomy describes how much freedom an AI agent has to decide and act without asking a person for every instruction. It is not an on-or-off property. The same agent can be autonomous in one part of a task and tightly controlled in another.

Autonomy has several dimensions

An agent’s autonomy depends on what it may control:

• Planning autonomy: Can it break a goal into its own steps?
• Tool autonomy: Can it choose which tools to use?
• Execution autonomy: Can it perform actions without approval?
• Recovery autonomy: Can it retry or change strategy after failure?
• Stopping autonomy: Can it decide when the task is complete?

For example, an assistant may independently research travel options but require approval before booking or paying. It has high autonomy in information gathering and low autonomy in financial action.

Autonomy comes from the surrounding system

A language model does not become autonomous merely because it can generate plans. The system must give it an agent loop, tools, permissions, state, and rules for continuing. The broader relationship between these components is covered in What Is an AI Agent?.

Permissions define what the agent can do. Instructions define what it should do. Guardrails, approvals, budgets, and step limits define when it must stop or escalate.

More autonomy creates more responsibility

Greater autonomy can reduce manual work, especially for long tasks with changing conditions. It also increases risk. An agent may misunderstand a goal, select the wrong tool, repeat an irreversible action, or continue with outdated state.

This is why useful autonomy should be bounded. High-impact actions such as sending money, deleting data, publishing content, or changing access rights often need deterministic validation or human approval.

Autonomy versus automation

Automation executes predefined rules. Autonomy involves choosing among possible actions based on a goal and current observations. A fixed script that sends the same report every morning is automated, not meaningfully autonomous.

The practical question is not “Is this agent autonomous?” It is “Which decisions and actions may this agent control, under what limits, and with what evidence?” That framing makes autonomy testable and governable.