Every vendor demonstration has the moment where the agent is asked something hard and the screen goes quiet while it reasons. You are meant to be impressed by the pause. Ask instead what it costs, and who decided this question deserved it.
Your best people do not think equally hard about everything. The finance director gives a renewal floor a minute and the annual plan a week, and that allocation is what makes a senior person affordable. The same allocation can be written into settings you can read: four of ten specialists given a model that reasons at length, and even those four held at the low setting. That is how the Cabinet is set, ArkOne’s reference design for an executive agent, and why the room thinks harder about fewer questions.
Two dials on a specialist’s thinking
A specialist here, one of ten Seats at the Table, is a name, a model and a set of instructions, as the sixth paper laid out. The model is where the thinking happens, and two dials sit on it.
The first is which model. Some models work through a problem in steps before answering, writing out their reasoning and checking it; they are slower and dearer per answer, and better on genuinely hard questions. The reference design gives such a model to four of the ten Seats and an ordinary model to the other six. Deciding which Seat a message belongs to needs no deliberation. A floor price defended against a client who negotiates for a living is another matter.
The second is how hard. A reasoning model can be told how much effort to spend before answering, from a low setting upwards, and more effort buys more time and a larger bill. The reference design holds its four deep-reasoning Seats at the low setting, and the Register, the one page that states every setting of the design, records the choice with the condition attached to it. That condition is a position, not a result. Every setting is put through the Rehearsal, the standing examination in which scored scenarios are answered by the room and marked by a judge model that is not the one being examined, and the rule is that a higher effort setting is bought once the Rehearsal can show what it buys. Its limit rides with it: the marker is another model, the scenarios are the ones somebody thought to write, and an examination that never asked a question cannot report on it. Paying more for a difference nobody has demonstrated is what the design declines to do.
Two honesties belong here. An unmeasured gain may still be real; the burden is only that someone show it before the money is spent, and the setting is then a line a person can change tonight, as the seventh paper described. The second is the cost of the choice: at a low setting a hard question may come back shallower. The remedy is a sharper question from the Chair, the one voice that runs the meeting, or a second consultation, both cheaper than paying the higher setting on every deep answer.
The same question, twice
Suppose Finance is one of the four deep-reasoning Seats, and the question is the one from the renewal running through this series: where does the floor price fall for the invented client asking to keep last year’s rate? Ask it twice, once at the low setting and once higher. The times below are modelled from the Register’s figure of roughly three times the time, illustrative rather than measured.
| What you would compare | The low setting, as the design runs it | A higher setting |
|---|---|---|
| Time to reply | About half a minute, modelled | About a minute and a half, roughly three times as long |
| The floor price returned | The floor computed from the terms supplied | The identical floor |
| The reasoning | Five sentences: margin, delivery cost since the route change, precedent risk | Fourteen sentences reaching the same three points |
| The Rehearsal’s mark | The same mark, in this modelled run | The same mark, in this modelled run |
| The bill for the reply | One unit of thinking | Roughly three units, for the same figure |
Exhibit 2. Illustrative. The same floor-price question answered at two effort settings; the figure is the same, the time is not.
Both columns were handed identical inputs and both ended with the Chair holding the list price. They differ in one row that matters to you, the time, and one that does not, the length of the reasoning. Multiply the extra minute by every deep consultation in a working day, across four Seats, and the higher setting has bought a slower room and a larger invoice for advice nobody has shown to be better.
The design withholds thinking in three places: reasoning models for four Seats and not ten, those four held at low, and, as the fifth paper showed, a Chair that consults the Seats a question needs rather than the roster. Each is a decision made once, by a person, recorded and open to change, rather than one the model makes afresh with your money. A scenario will exist somewhere on which the higher setting finds what the low one missed; when the Rehearsal shows it, the setting moves and the Register changes with it.
What this arms you to ask
Ask the vendor which model answers which question, and who decided.
A good answer is a table: each specialist, the model behind it, how hard it may think, the person who set it, and the test that would have to come back different before the setting moved. That vendor can also show you the same question answered at two settings.
A weaker answer is that the agent runs on the best model available and reasons deeply about every question. Read that as a bill. One model at one setting prices the trivial question and the hard one alike, and nobody decided which deserved the pause, the choice the eighteenth paper hands back; your money goes at a rate set by whoever chose the default. Ask the follow-up: show me the same question at two settings and the difference in the answer. Then ask who may change the setting, and whether they work for you.
Next paper: What the Chair reads before every meeting, in the order that decides your bill.
Asked plainly
Do AI agents need a reasoning model for every question?
No. Models that work through a problem in steps before answering are slower and dearer per answer, and on routine questions they produce the same answer as an ordinary model. The reference design gives such a model to four of its ten specialists and an ordinary model to the other six, so the expensive thinking is spent only where a question is likely to need it.
What is the effort setting on an AI reasoning model?
How much thinking the model is allowed to do before it answers, from a low setting upwards. More effort means more time and a larger bill for the same question. The reference design holds its reasoning specialists at the low setting because its scored examination could not tell the answers at higher settings apart from the answers at low.
Who decides which AI model answers which question?
In a well-designed system, a person, once, per specialist, and the choice is recorded where it can be read and changed. In a system that runs one model at one setting for everything, nobody decided: every question costs the same, the trivial and the hard alike, at a rate set by whoever chose the default.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.