Put a hard question to your leadership team and watch how the answer is assembled. You do not ask Finance, wait for a reply, then ask Counsel, wait again, then ask Operations. You ask all three at once, and the answer takes as long as the slowest of them needs. An agent that consults its specialists one after another has given up that advantage before it started.
The Table is the part that lets one voice ask many. It belongs to the Cabinet, ArkOne’s reference design for an executive agent, where the Chair, the single voice that runs the meeting, can put a question to several specialists in one round and hear them answer together. The sentence to carry away: the Chair asks everyone it needs in one breath, and the ceiling on how many is the size of the roster.
Why the room answers together
A specialist in this design is one of ten Seats at the Table. Taken alphabetically, so that no order is implied: Board, Counsel, Finance, Marketing, Operations, People, Product, Strategy, Talent and Triage. When the Chair wants a view, it writes a request naming a Seat and the question, in the way the third paper described for any tool. What the Table adds is that one request may name several Seats. The room sends each its question at the same moment, waits for all of them, and hands the replies to the Chair as one bundle of material.
The consequence is in the clock. A round with three specialists in it lasts as long as the slowest specialist, not as long as all three added together.
Parallel is not always right, and the Chair is not made to use it. When the second question depends on the first answer, the Chair asks in turn. In the renewal traced through this series, the invented client asking to keep last year’s price, Finance was consulted alone in the second round, because what to ask Marketing and Operations depended on where the floor price fell; those two were then asked together in the third. That judgement is the Chair’s, each round.
The dial on the Table is the fan-out cap: the most Seats one request may name. The reference design sets it at the size of the roster, so with ten Seats, ten may be asked, and the Register says so. Two things follow, and the second is a limit you should hear plainly. A request that names more Seats than the cap allows does not fail quietly; every Seat over the cap receives an explicit skip message, so the Chair knows who was not asked. And a cap set at the roster is a ceiling, not a brake. Nothing in the Table stops the Chair consulting all ten every time. What stops it is the same thing that stops you: each consultation is a reply that must be paid for and read, and the fifteen rounds allowed by the first paper bound how many rounds there are to spend.
One round, three Seats
Suppose a different question from the renewal, with the company invented for the exercise. A competitor has cut its prices by a fifth in the company’s main market, and the chief executive asks the room what to do this quarter. Round one fetches the price list and the open pipeline from the Library, the room’s own shelves of documents. Round two is the one this paper is about.
| Seat asked, in the same round | The question the Chair wrote | What came back |
|---|---|---|
| Strategy | Is this a permanent move or a quarter’s promotion, given what we hold about this competitor? | A promotion, on the evidence: their last two cuts were reversed within six months. Hold |
| Finance | If we matched the cut across the top twenty accounts, where does margin land? | Below the floor on eleven of them; a match is unaffordable at current volume |
| Marketing | Which accounts are most likely to be approached, and what do they value beyond price? | Six named accounts; four of them cite delivery reliability above price in this year’s survey |
Three Seats of ten were asked, and three replied in one round. If each took half a minute, the round took half a minute; asked in turn, the same three answers would have taken a minute and a half. The dial in the exhibit shows the round as the design sees it: ten on the roster, three used.
Why not all ten? Board had nothing to add to a quarter’s pricing question, and Talent less. Each reply that comes back is material the Chair must read before it can speak, and a reply that adds nothing still costs a consultation and lengthens everything the Chair reads afterwards, which is the subject of the ninth paper. The eight of the title is what a genuinely wide question looks like: the annual plan, where every Seat but Talent and Triage might be asked at once. The ceiling is the roster, which makes that legal. It does not make it cheap.
Fifteen is the cap on the Register, not a target. Press run and count them.
A meeting has not been run yet. Press run to watch one count, fan out to the specialists, come back, and stop.
What this arms you to ask
Ask the vendor how many specialists the agent can consult at once, and what stops it consulting all of them every time.
A good answer has two halves. The first is a number, where it is set, and what a specialist over the cap receives; the vendor can show a transcript with several specialist requests carrying the same timestamp. The second half is the honest one: the cap is a ceiling, and what keeps the room from asking everyone is the orchestrator’s judgement and the bill, both of which you can read afterwards.
Listen instead for a sentence with no roster behind it: the agent draws on whichever expertise the question needs. Suppose the thing behind that sentence is one model with one set of instructions, considering the finance angle, the legal angle and the operations angle inside a single reply. Then there is nothing to consult in parallel because there is nobody to consult, and the specialists are headings in one document. Ask the question that separates the two: show me two specialist replies that arrived in the same round.
Next paper: What a specialist actually is: thirty lines and a set of instructions.
Asked plainly
Can an AI agent consult several specialists at the same time?
In a design built as a room of specialists, yes. The orchestrator writes one request naming several specialists, each is asked at the same moment, and the replies come back together. A round then takes as long as the slowest reply rather than the sum of all of them. Where the second question depends on the first answer, the orchestrator asks in turn instead.
What is a fan-out cap in an AI agent?
The most specialists one request may name. In the reference design it is set at the size of the roster, so all ten may be asked at once. A specialist over the cap receives an explicit skip message rather than being dropped in silence, so the orchestrator knows who was not consulted.
What stops an AI agent asking every specialist every time?
Not the cap, which is a ceiling. What stops it is that every reply must be paid for and read before the agent can answer, and the orchestrator's own judgement about which specialists a question needs. Both are visible afterwards in the transcript, which is where to check.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.