Before it builds me a board pack, do I see the plan or only the result?
Ask for something long. Every demonstration has a showpiece, and it usually finishes in under a minute because it was chosen to. Instead give the agent a job with a dozen questions inside it, in a fixed order, some depending on the answers to others. A board pack. A supplier review. Then sit on your hands and watch the first ten seconds, because those ten seconds decide whether you can ever change your mind cheaply.
The cheap look, and where it sits
A multi-step deliverable is a sequence of consultations, and the consultations are where the money goes. Writing down the sequence is different in kind: it asks nobody for anything, so it costs one inexpensive turn.
In the Cabinet, ArkOne’s reference design for an executive agent, such deliverables sit on the Order Paper, the named multi-step pieces of work the room knows how to run. The rule on the Register is the plan, first. Before a single step runs, the whole sequence streams to you: which figures it will fetch, which specialists it will ask, where it will pause for your word. Only then does step one begin. The limit belongs in the same sentence. A plan is a forecast, so a run can discover it needs something the plan did not name, and the revision streams before the revised step runs; what the earlier steps cost has already been spent.
The same job, three chances to speak
Suppose you ask for a quarterly board pack in the demonstration, invented for the exercise, and you interrupt at the earliest moment the design allows.
| Where you get to speak | What is on the screen | What it had spent by then |
|---|---|---|
| Ten seconds in | A numbered sequence naming every step and every specialist it will consult | One cheap turn, nobody consulted |
| Ten seconds in, without the rule | A progress line, or the first figures arriving | Whatever step one cost, and it is running |
| Partway through | A named pause, waiting for your word before the next step | The steps before the pause, and no more |
| At the end | A finished pack and a record of how it was made | All of it |
Exhibit 1. Illustrative. A long job asked for in a demonstration, and what had been spent each time you could speak.
Compare the first two rows, which are the same ten seconds. In one you can strike out a step for the price of a single turn. In the other your correction arrives with the first step already bought, and if that step was the expensive consultation, you have paid for the thing you were about to remove.
The fourth row is where most conversations about this end up, and it is not nothing. A record of how a pack was made is worth having at the audit, and you should ask for it. It is simply not a control, because everything it describes has happened.
Stop it, in front of me
The good answer is a demonstration rather than a sentence. The steps appear named before any of them runs, the vendor stops the run there to prove it can be stopped, and somebody can say what the run had spent at that instant.
Where that has not been built, the reply usually points at visibility of a different kind: that the agent shows its reasoning as it works, or that every step is in the log afterwards. Both describe something real, and both arrive after the spending starts. The narrower follow-up gets past the ambiguity: at what moment in this run could I have stopped it, and what had it already spent when that moment arrived? A vendor who has thought about it answers in two numbers. A vendor who has not will describe the log again.
Then ask who may define a deliverable. If a board pack is something only the vendor can add to the product, the shape of your work is bounded by their release schedule. If you can name a new sequence yourself, in plain steps, and watch its plan stream the first time it runs, the design has room for how your company actually works. That second question tells you more about the next three years than the first one does.
What ten seconds buys
The whole argument is about when you meet the plan rather than whether one exists. Every system has a plan; the question is whether it reaches you while changing it is still free. The rule, its cost and its honest limit are worked through in the paper on the Order Paper. What the room does with your reply at one of those pauses is the question before this one.
Asked plainly
Should an AI agent show its plan before it starts work?
For anything with several steps, yes, and the reason is arithmetic rather than courtesy. Writing the plan consults nobody and costs one cheap turn, while the steps are where the spending happens. A wrong assumption caught in the plan costs that one turn; the same assumption caught in the output has been paid for in full.
Can I stop an AI agent partway through a long job?
Only where the design gives you a moment to speak into. A run that shows its plan and waits, and pauses again at named approval steps, has such moments by construction. A run that works first and narrates as it goes may be stoppable, but every stop arrives after the step it would have prevented.
Is a log of what an AI agent did the same as seeing its plan?
No, and the difference is when you see it. A log is written as steps complete, so it can only tell you what has already been paid for. A plan is written before any step runs and can be corrected while the correction is still free. Both are useful; only one of them saves anything.