From text to action

An assistant suggesting a reply and an agent sending a message have different responsibilities. Before introducing an agent, list the tools it can use, the data it can read and the operations it can initiate. A task description needs effective controls over those capabilities.

Permissions belong outside the model

Asking a model not to do something is different from preventing the action. Permissions should be enforced by the tools and systems that execute operations. An agent reading documents could have read-only access to a specific collection without permission to delete it or share it externally.

Approval where the risk changes

Not every step needs human intervention. Distinguish ordinary tasks from operations with broader consequences, such as a production change or sending data to a new recipient. These need clear approval steps and a record of the decisions.

The AgIPS starting point

AgIPS proposes connecting action controls to the context of the task. A trial should check adherence to permissions as well as the ability to complete legitimate work. Excessive blocking can make a system unusable, while insufficient blocking can leave it exposed.

Further reading