You type one line: board pack for Thursday, same shape as last quarter, with the pricing risk called out. What appears in the next ten seconds tells you more about the agent than any demonstration did.
What appears first, in the Cabinet, ArkOne’s reference design for an executive agent, is a plan. Seven steps, each named, each saying which specialists it will consult and where it will stop for your word. No figure has been fetched and no specialist asked. You read it in the time a kettle takes to boil, and if step four is wrong you say so while the room has spent one turn instead of seven steps.
A deliverable is a sequence, and the sequence is shown first
Most of what you ask an agent is a question, answered in one meeting. A board pack is different in kind: a dozen questions in a fixed order, some depending on the answers to others, with a person’s approval needed partway through. The Cabinet keeps such things on the Order Paper, the set of named, multi-step deliverables the room knows how to run: board packs, investor updates, monthly reviews, risk registers, and others you define yourself, without code, by naming the steps.
Each deliverable is run by the Chair, the one voice that speaks for the whole room, and each step may consult the Seats, the ten specialists around the Table. The rule that matters is on the Register, the page that records every setting of the design: the plan, first. Before any step runs, the Chair writes the whole plan and streams it to you. Which figures it will gather and from where. Which Seats it will ask, and what it will ask them. Where it will pause at a gate, an approval step that waits for a person. Only when the plan has finished streaming does the first step begin, and your word at any point stops it.
The reason is money and the order it is spent in. A plan is not free: it is one turn of the Chair, billed like every turn, as the ninth paper sets out. It is merely cheap, because it consults nobody, and the steps are where the money goes. Planning first puts the cheap look at the front, where a wrong assumption costs a single turn. A design that works first and narrates as it goes has already paid for step one by the time you can see step one was wrong.
Two limits belong beside that. A plan is a forecast, and a run can discover it needs a Seat the plan did not name; the revision is streamed before the revised step runs, but what the earlier steps cost has been paid. And the stop is your word, in a message: a plan that streams past an empty chair has been shown to nobody.
The Thursday board pack, traced
Suppose the invented company of these papers, the one whose client asked to renew at last year’s price, wants a board pack by Thursday. Here is what reaches the reader at each moment, and what the room has spent by then.
| Moment | What streams to you | What the room has spent |
|---|---|---|
| The request | Your one line, echoed | Nothing yet |
| The plan | Seven steps: gather the quarter’s figures; ask Finance to size the pricing risk; ask Operations for delivery cost; draft the narrative in last quarter’s shape; assemble the pack; pause at a gate for your approval; distribute through the Door | One turn of the Chair, no specialists |
| Your word | You reply: leave Operations out, the board saw it last month | Still one turn |
| The revised plan | Six steps, Operations removed, streamed again | A second turn, no specialists |
| Steps one to five | Figures fetched, Finance consulted, narrative drafted, pack assembled; each written to the Record as it completes | Five steps of the loop, with their consultations |
| The gate | The run pauses: the pack is ready; approve, adjust, or hold. It records your answer and stops, and a person restarts it | Nothing further |
| Distribution | The pack leaves through the Door to the board list, each recipient checked against the Directory | One send |
Exhibit 2. Illustrative. A board pack requested on a Tuesday: what streams to the reader at each moment, and what the room has spent by then.
Two rows are easy to miss. In the third, your correction arrived when the room had spent one turn of the Chair and consulted nobody, so rewriting the plan cost a second. Had the design worked first and explained afterwards, the Operations consultation would already be on the bill and in the draft. The saving comes from the arithmetic: two turns of the Chair against seven steps of consultations.
In the sixth, the gate does not wait for you and then carry on. It records your decision and stops, by the same rule the Signature follows, the step that waits for a person elsewhere in the design. The fourth paper explained why an approval must not become an assumption, and the twenty-second takes up an approval that is not quite one.
What this arms you to ask
Before it builds me a board pack, do I see the plan or only the result?
A good answer is a demonstration in which the steps appear on screen, named, before any of them runs, and the vendor stops the run there to prove it can be stopped. That vendor has separated planning from working. The answer that should worry you names no mechanism: the agent shows its reasoning as it goes, or every step is in the log afterwards. Both describe a plan you meet after the spending has begun. Ask a narrower question: at what moment in this run could I have stopped it, and what had it already spent when that moment arrived?
Then ask who defines the deliverables. If a board pack is something only the vendor can add, your work is as specific as their roadmap allows. 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 the way your company works.
Next paper: The gate that waits: what happens when your approval is vague.
Asked plainly
Can an AI agent produce a board pack or an investor update?
Yes, if the agent treats it as a named sequence of steps rather than a single question: gather the figures, consult the specialists, draft the narrative, assemble, pause for approval, distribute. In the reference design these sequences are called deliverables, and each one shows its plan before it starts on the work.
Why should an AI agent show its plan before it starts?
Because a plan is cheap and the work is not. Writing the plan is one turn of the agent with no specialists consulted; each step that follows consults specialists and fetches documents, and every one of those is billed. A reader who sees the plan first can correct a wrong assumption for the price of that single turn.
Can I stop an AI agent partway through a long task?
In the reference design, yes, at the plan and at every approval gate the deliverable contains. Two limits are worth knowing: a plan is a forecast, and a run that has to revise it has already spent what the earlier steps cost; and stopping is a person's word, so a reader who does not watch the plan stream by has not stopped anything.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.