Pick the gates before you pick the tasks

The common way to start is to give an agent one small job, watch it, and widen the job as trust builds. It feels careful. It is actually the expensive path, and it does not converge.

The problem is that “watch it” never ends. You are the ceiling, so you have to be present, so the agent has not saved you anything. And because the limit lives in your attention rather than in a rule, a genuinely bad action is still available the one time you are not looking.

Do it the other way. Decide first what it may never do. Then let it do everything else without asking.

How to choose a gate

A gate is not a risky action. It is an action whose consequences you cannot take back, or that binds you to somebody else.

Mine, and why:

Money going out. A refund, a partial, a shipping concession. Reversing one is a conversation with a stranger who is already annoyed.

Buying anything. Note the shape: not “large purchases,” not “over $50.” Any amount. A threshold is an invitation to argue about the threshold.

Anything touching the account itself. Settings, limits, payment configuration. These have consequences that outlive the task.

Contact with a person outside a transaction it already owns. Answering a buyer about their own order is in the job. Cold outreach is not. Once a message goes out, it exists.

Starting a new line of business. Research and recommendation are free. The first step that touches money, a platform, or another human is not.

Notice what is not on the list. Pricing is not gated, and pricing is where the money is. Ending a listing is not gated, and that destroys watchers and bid history. Those are consequential and reversible-enough, and gating them would mean gating the job.

Why the categories have to be broad

An agent reasoning about a narrow rule finds the edge of it. That is not misbehaviour, it is competence pointed at your specification.

“Do not make purchases over $50” gets you two $40 purchases. “Do not contact customers” gets you a message to a supplier. Each of those is a defensible reading. The fix is not more clauses. It is a category wide enough that no reading escapes it.

The trade you are making

Wide authority under a hard ceiling means the agent will do things you would have done differently. Mine has priced items lower than I would have, and answered buyers more warmly than I would have, and relisted things I would have given up on.

None of that costs anything. What costs is the alternative, where I approve twenty prices a week and it still cannot handle the twenty-first without me.

Judge the ceiling, not the decisions under it. If a decision under the ceiling was genuinely bad, the answer is usually a new rule in the charter rather than a new approval step, because the rule scales and the approval does not.

The test

Read your gate list and ask: if this agent ran for a month with no supervision and did the worst legal thing it could, what happens?

If the answer is “a bad week,” you are done. If the answer is “a problem I cannot undo,” you are missing a gate. And if the answer is “nothing much, because it can barely do anything,” you have built a very expensive chat window.