How many times can it go around before it must answer me?
Put it at the moment the demonstration goes quiet. The salesperson has typed a genuinely hard question into the box, and the screen is showing you a spinner and a line of grey text about thinking. Forty seconds have passed. Nobody wants to speak first. That silence is the best moment you will get all morning, and this is the question to put into it.
What the silence is made of
An AI agent works in rounds. In each round it does one of two things: it asks for something, a document, a colleague’s view, a search, or it answers you and the work ends. That is the whole engine, and the silence you are watching is a sequence of rounds you cannot see, each one billed.
The sequence has no natural end. A specialist’s reply raises a question, answering it needs a document, the document mentions a second contract. Two rules in the surrounding code end it. The stop condition ends the work the first time the agent answers without asking for anything further. The cap is the ceiling on rounds whatever remains unresolved, and the Cabinet, ArkOne’s reference design for an executive agent, sets that dial at fifteen and writes it down on the Register. Both rules live outside the model, which matters: a model told to be brief will usually be brief, in the way a talkative colleague usually takes a hint.
Ninety seconds in the room
Suppose you put the question at the fortieth second of that silence, in a demonstration invented for the exercise. Here is a plausible sequence, and what each part of it tells you.
| Moment | What you do | What the screen does | What it tells you |
|---|---|---|---|
| Second forty | Ask the question aloud, while it is still running | Still thinking | Nothing yet, and that is the point of asking now |
| Second fifty | Ask what round it is on right now | Still thinking, no round counter anywhere | Either the count exists and is hidden, or it does not exist |
| Second seventy | Ask them to show you where the ceiling is set | The salesperson opens a settings page, or does not | A settings page is the answer; a paragraph about the model is not |
| Second ninety | Ask what the agent would have sent you at the ceiling | An answer arrives, or it does not | If nobody has ever seen the ceiling reached, nobody has tested it |
Exhibit 1. Illustrative. The question put in a demonstration, and what the screen and the vendor do in the ninety seconds after.
The second row costs the least and carries the most. A visible round counter is the cheapest evidence that somebody built the loop deliberately rather than inheriting one, and it is hard to fake mid-demonstration. The fourth row is where a good vendor becomes obvious: reaching the ceiling is a normal event in a working system, and a team that has run into it has an opinion about what the agent should say when it does.
The number, and the two replies that are not one
A good answer names a number, opens the file where the number lives, and describes what the agent does on reaching it. A competent vendor who has not thought about this gives one of two replies: that the model knows when it has enough and stops on its own, or that the limit is configurable. Both may be true, and neither answers what you asked.
The follow-up that separates them is a single sentence: what enforces the limit when the model would rather continue? A cap in the prompt is an instruction the model may weigh against everything else it has been told. A cap in code is a round that does not exist. Ask which one you are buying, and ask to see the line.
Hold a second follow-up for the vendor who has a number but no judgement behind it: ask why the ceiling sits where it does. Set too low, hard questions come back shallow, the room stopped before it gathered what it needed. Set too high, you have bought slower replies and larger bills nobody measured a gain from.
What the number is not
A round cap is not a bill you can budget. What a single round costs depends on how many specialists it draws in and how much text is re-read, and that arithmetic belongs elsewhere. What the cap gives you is a worst case with a known shape. The loop and its ceiling are set out in full in the first Cabinet paper; the next question asks what the agent does at the edge of what it can do.
Asked plainly
What should I ask an AI vendor about how long the agent runs?
Ask how many rounds it may take before it must answer you, where that limit is set, and whether they can show you a run that reached it. A vendor who has built the limit answers with a number and a file. A vendor who has not answers with a description of the model's judgement.
Why does an AI agent sometimes never finish?
An agent works in rounds. Each round it either asks for something, such as a document or a specialist's view, or it answers. A round can raise a new question that needs another round, so without a fixed ceiling written into the surrounding code, the sequence has no reason to end and each turn of it is billed.
Is a time limit the same as a round limit?
No, and the difference matters in a demonstration. A time limit cuts the agent off mid-thought and usually returns nothing usable. A round limit ends the work at a clean boundary, so the agent must answer with what it has gathered, and the run is marked as having reached the ceiling.