An agentic action is an operation an AI assistant performs beyond generating text, such as querying an API, retrieving a file, changing a record, or triggering a workflow. It is carried out through a connected tool or system, usually in response to the assistant’s plan or a user request.
These actions matter because they can access sensitive data or cause real-world changes. In a retrieval-augmented system, permission to answer a question does not automatically grant permission to call tools or modify resources. Defenders should authorize each action separately, apply least privilege, verify inputs and user identity, and log what was accessed or changed. These controls help prevent an assistant from turning excessive access or unsafe instructions into unauthorized operations.