How Much Should An AI Agent Decide On Its Own?

The short version

"How autonomous should it be" is usually asked as one question. It is really two: how much routine work runs without asking, and who authorizes spend. The second answer should not change. The first should, as you learn how the agent handles your account.

What you are actually setting

The permission level controls how often the agent stops to confirm before acting.

Level	Behavior
Ask permission	Asks you to approve each creative and each ad setup
Auto approve	Approves creatives and ad setup for you, and checks in when direction is unclear
Autopilot	Runs on its own and sets its own direction

Auto approve is the default and the right setting for most work.

Spend and deletion pause for approval at every level, Autopilot included. This is the part that gets misread. Raising the level changes how much routine work runs unattended. It does not delegate your budget, and there is no setting that does.

Calibrate rather than choose

The useful approach is not picking a level once. It is starting low on a new account and moving up as you accumulate evidence.

Start at Ask permission on an account you have not run this way before. You are not protecting yourself from catastrophe — spend is gated regardless — you are learning whether the agent's creative and setup judgment matches yours.
Move to Auto approve once the approvals stop surprising you. If you have approved fifteen creatives without changing anything, the gate has become a formality and formalities erode attention.
Use Autopilot for work you have already watched it do well. It suits repeated, well-understood workflows rather than novel ones.

Moving up too fast costs you the calibration. Staying too low costs you the attention you will need for the approvals that matter.

What triggers a confirmation

When the agent asks to act, the request is labeled with the category of permission involved: campaigns and ads, audiences, image and video generation, creatives, brand assets, scheduled tasks, experiments, and brand profile.

Read-only actions run without confirmation. Nothing is gained by approving a report being pulled, and a queue of trivial approvals is precisely how people stop reading the ones that matter.

Who does the work

The work is split across specialists rather than handled by one general agent. A director decides what to run and which specialist executes each step; the others cover market research, creative direction, media buying, and analysis.

You do not route work yourself. Stating the goal is normally enough.

Depth is a separate dial

Sessions can run at different reasoning depths, from Instant through to Comprehensive. Deeper settings take longer and cost more, and are worth it for planning and diagnosis rather than for routine execution.

This is independent of the permission level. A shallow-but-autonomous session and a deep-but-gated one are both reasonable configurations for different jobs.

Scheduled runs

Work can be scheduled to run unattended. A genuinely unattended run locks the conversation for its duration, since there is nobody there to answer a question. Anything that spends still waits.

Three habits that make it work

State the goal and the constraint together. "Improve ROAS" is weaker than "improve ROAS without raising spend above the current level."
Answer the direction check-ins. When the agent asks about direction, that is the cheapest moment to correct a plan. Ignoring it and correcting the output later is the expensive path.
Read the spend approvals properly. They are deliberately the few that reach you.

Practical rule

Start at Ask permission, move up when the approvals stop teaching you anything, and never treat the level as a decision about your budget. That decision is not on this control.