Action

An action is something an AI agent does to make progress toward a goal. It can be an external operation, such as calling an API, or a communicative step, such as asking the user for missing information.

Actions turn decisions into effects

The model may decide that a customer’s order status must be checked. That decision becomes an action only when the system calls the relevant order tool with specific arguments.

Typical actions include:

• Searching for information.
• Reading or writing a file.
• Querying or updating a database.
• Running code or tests.
• Sending a message.
• Requesting human approval.
• Returning a final answer.

Tool Use in AI Agents explains how an agent selects tools, prepares inputs, executes calls, and handles their results.

Every action should produce an observation

After an action runs, the environment returns information about what happened. That observation may contain a result, confirmation, warning, or error.

For example, a scheduling agent may call a calendar tool to create an event. The response should confirm the event ID and time. If the call times out, the agent must not blindly retry because the event may already exist.

This is where idempotency matters: a safely repeatable operation should not create unintended duplicates. When actions are not idempotent, the agent may need to check the environment before retrying.

Action is not the same as reasoning

Reasoning helps choose what to do. An action changes or queries the environment. A model can reason about sending an email without actually sending one.

Keeping this boundary clear is essential for safety. The system can allow broad reasoning while restricting external actions through permissions, schemas, validation, and approval gates.

Actions need boundaries

The impact of an action determines how much control it needs. Reading public information may require little oversight. Deleting data, spending money, publishing content, or changing access rights should usually require stronger checks.

A well-designed agent exposes the smallest set of clear actions needed for the task. Narrow tools reduce ambiguity and make logs, testing, and authorization easier than one overly powerful generic action.