Your phone shows one line from the room: the investor update is assembled and addressed to twelve people; approve, adjust or hold. You are between a car and a dinner. You type, looks fine, maybe tighten the second paragraph, and put the phone away.
Read that back. You approved it. You also asked for a change, and did not say what the change should be. A colleague would know to check. Whether the agent does is what the Cabinet, ArkOne’s reference design for an executive agent, settles at a gate: the step in a multi-step piece of work that waits for a person before the run continues.
Two rules at the gate
The investor update is a deliverable on the Order Paper, the set of named multi-step pieces of work the room runs to a plan, described in the twenty-first paper. Partway through stands the gate, the same fixture as the Signature, the step that waits for a person anywhere in the design. Two rules govern what happens when your reply arrives, and both are on the Register, the one page that states every setting.
The first is the resume rule. When you answer, the run records your decision and stops, and a person restarts it. Even an unmistakable yes does not carry the run forward on its own; the yes is written to the Record, the ledger of every action and its reason, and the next step waits for a hand, which can be yours one tap later or a colleague’s on Wednesday. The approval and the action never happen in the same breath with nobody between them. A slow answer means a delayed update, and silence means nothing leaves.
The second is the confidence rule. Before recording your reply as a decision, the design asks how sure it is of what you meant and puts a number on it. An approval reply it is less than 85 percent confident it has understood is asked again, never guessed. Asked again means a short message back naming the readings it could not choose between: send as it stands, or hold while the second paragraph is redrafted? Your next reply is measured the same way.
The limit sits inside the rule. The confidence is the model’s estimate of its own reading, so a reply it misreads with conviction passes the threshold. That is why the two rules exist together. The confidence rule catches the replies the design knows are unclear; the resume rule means even a misread reply cannot send anything, because sending waits for a person who sees what was recorded. The Directory, the roster of people the room knows, adds a third check that needs no number: a reply from anyone whose approval rights do not cover the update is no approval at all, and a reply from a stranger is dropped unanswered.
Four replies, one gate
Suppose the investor update of the invented company sits at its gate on a Tuesday evening, addressed to a dozen investors, and four versions of you reply.
| Your reply | What the design reads | Its confidence | What happens |
|---|---|---|---|
| Yes, send it | A clear approval, no change asked for | Well above the line | The approval is recorded and the run stops. A person restarts it; the update leaves through the Door, the rules that decide what may leave |
| Looks fine, maybe tighten the second paragraph | An approval and an unspecified change, at once | Below the line | Nothing is recorded as a decision. One message comes back: send as it stands, or hold while the paragraph is redrafted? |
| Not yet | A clear hold | Well above the line | The hold is recorded and the run stops. Nothing is addressed, nothing leaves, and the pack waits |
| No reply | Nothing | Nothing to measure | The run waits at the gate. Nothing is assumed, and the morning brief shows it waiting for your word |
Exhibit 2. Illustrative. Four replies to the same approval request: a clear yes, a vague yes, a no, and silence, and what the room does with each.
The second row is the one to dwell on. A design without the confidence rule would do something with that message, and whichever thing it did would be defensible afterwards: it sent, because you said looks fine; or it held, because you asked for a change. Defensible is the problem. You would learn which reading it took when twelve investors did or did not receive the update, and the transcript would show your own words as the authority. The reference design hands the ambiguity back to the only person who can resolve it, with the two readings named, so resolving it costs you a word.
The first row matters for the opposite reason. Nothing was ambiguous, and still the run stopped. The design treats a restart as a deliberate act, because a yes that flows straight into a send is a gate in name only. The fourth paper set out why actions that leave the room are a kind of their own; a gate that continues by itself would let a well-worded message from the wrong phone move an irreversible action along unnoticed.
The fourth row is where the design is more conservative than a colleague, who might send the update anyway by nine, knowing your habits. The room waits, built so that the absence of a decision is never read as one.
What this arms you to ask
When it pauses for my approval and my answer is vague, what does it do?
A good answer is a live demonstration. The vendor types a reply of the looks-fine-but kind, the agent asks again with the two readings named, and the vendor points to where the threshold is set and to a record of a reply that fell below it. Two answers put the burden back on the model: that it understands natural language well, and that it asks if something is unclear. Ask what decides unclear. If it is a judgement the model makes freshly each time, with no number and no setting, the line between an approval and a guess moves with every message. Ask, too, what happens after a clear yes. If the update leaves in the same motion, ask who was between your word and the send.
Next paper: What lands on your desk at eight, and what it leaves out on purpose.
Asked plainly
What does an AI agent do when my approval is ambiguous?
It depends on the design. Where nothing measures the reading, an agent proceeds on its best guess at your words, and the guess is invisible until the result arrives. The reference design measures how confident it is that it understood the reply, and when that confidence falls below a fixed threshold it asks again, in a message naming the two readings it could not choose between.
After I approve, does the AI agent finish the job on its own?
Not in the reference design. An approval is written down and the run stops; a person restarts it, which can be the very next thing that person does. The rule exists so that an approval can never quietly become an assumption, and so that a slow or absent answer means a delayed action rather than an assumed one.
Who is allowed to approve an AI agent's actions?
Only a person the agent knows, whose named approval rights cover the thing being approved. In the reference design a reply from anyone else is not an approval at all, however clearly worded, and a reply from an unknown sender is dropped without an answer.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.