ARKONE
← The Cabinet
Case note · silentThe Directory · the approved-contact list5 minute read

"I will loop Jordan in", said to Jordan

An agent that does not know who it is talking to will write about that person as though they were absent.

Silent. Your operations director asks the agent a question about a stalled order. The reply is competent, and it closes by offering to loop in the operations director, whose name is Jordan, so that Jordan can confirm the position. Jordan reads this. Jordan concludes that whatever the company has bought is not paying attention, and mentions it to nobody, because it is a small thing and complaining about it feels petty.

Every part of that exchange worked as built.

What went wrong

The agent knew who Jordan was. It had a roster with Jordan on it, a title, a department, an address. What it did not have was any indication that Jordan was the person it was writing to.

Those are two different pieces of information, and this is the whole of the mechanism. A roster is stored: it sits in a table and does not change between one message and the next. The identity of the current correspondent changes on every single request, so it has to be placed into what the agent reads each time. If no line of code does that placing, the agent is handed a question, a roster of colleagues, and no marker saying which chair it is speaking from.

Given that, offering to forward a message to the reader is not a lapse in reasoning. It is the correct output. The agent sees a matter that would benefit from the operations director’s confirmation, it sees an operations director on the roster, and it does the sensible thing.

The Cabinet, ArkOne’s reference design for an executive agent, treats this as a question of admission rather than instruction. The Directory, the roster of the real people the room knows and may address, is consulted before the meeting starts: a message from anyone not on the roster is dropped without a reply, and a message from someone on it arrives with that person’s entry attached, including what they are permitted to approve. Recognition happens at the door, and the answer travels with the request. The same gap in a different place produces an agent signing as the chief executive: there the question is whose name goes out, here it is whose name comes in. The twentieth paper sets out the roster; the second sets out why the room speaks with one voice, which only works if it knows whom it is speaking to. Both settings sit on the Register.

The record

Suppose a logistics firm of four hundred people, with an agent answering internal questions. The firm, the people and the exchange are invented for the exercise.

What the agent was given What it wrote Why
The question, and the roster ”I will loop Jordan in so operations can confirm” Jordan is the right person and nothing said Jordan was asking
The question, the roster, and the sender’s entry ”You will know better than the record does, so confirm the revised date” The reader is the operations director, so the ask goes to them

Exhibit 1. Illustrative. The same question, with the recipient's identity absent and present.

The second row is what one field changes. Same model, same roster, same question, same instructions. The only difference is that the sender’s identity arrived with the request instead of being left out of it.

What makes this worth a note of its own is the size of the class it belongs to. A missing recipient does not produce one recognisable bug you could name and fix. It produces a promise to forward a message to the reader, an explanation of the reader’s own remit back to them, an approval routed to somebody with no standing to give it while the person who does have that standing is on the line, and a stranger addressed with the familiarity of a colleague. Every one of those reads to the person on the other end as inattention, and none of them reads as a fault to whoever built the thing, because the output is fluent and the log shows a normal exchange.

The defenceThe sender’s identity is attached to the request, so the agent reads who it is speaking to.

What happens

The agent addresses the reader as the person they are, and never offers to forward a message to them.

What your company sees

The action was refused and the reason recorded.

Illustrative. One setting, switched. Its value is on the Register.

What it means for you

The reason this matters beyond the embarrassment is that the instinct to fix it is almost always wrong. The reply looked like a judgement error, so the repair looks like better wording: add a line to the instructions telling the agent to consider who it is addressing.

That cannot work. The agent was not misjudging the fact, it did not have the fact. An instruction tells a model what to do with information in front of it and can never supply information that is absent. Write the line and the behaviour persists, at which point people conclude the technology is not ready.

So the question is a plain one. When somebody sends your agent a message, what identifies the sender, and where does that identity appear in what the agent reads. A good answer names the point in the code where the two are joined, and says what happens when the sender is not recognised. If the answer is that the agent knows the team, ask whether it knows which of them is speaking.

Next: every case note, in order.

Asked plainly

Why does an AI agent write as if it does not know who sent the message?

Because knowing a list of people and knowing who is on the other end of the current conversation are two different pieces of information. The first is stored data. The second has to be put into the request every time the agent is asked something. If nothing does that, the agent has a roster of names and no idea which of them it is speaking to.

How does an AI agent know who it is talking to?

Whatever receives the message identifies the sender, usually by matching the address or handle it arrived from against a roster of known people, and then writes that person's identity into the material the agent reads before it answers. It is a field in the request. It is not something the agent can work out from the wording of the message.

Can you fix an AI agent's mistakes by improving its instructions?

Only mistakes of judgement. If the agent lacks a fact, no instruction supplies it: telling a model to consider who it is addressing does nothing when the answer is not in front of it. Missing information is fixed by putting the information into the request, and instructions are for what to do with it once it is there.

Talk it through before you decide

A discovery call, no deck: your situation, the parts of the room it touches, and what you would need to decide first.