Stack three decisions into one message and an agent answers the easy one properly and gestures at the other two. Interaction design for a running agent has more in common with air traffic control than with chat.
When I give an agent three decisions to make at once, I get back an answer that addresses one of them properly and gestures at the other two. It is not that it cannot reason about three things at once — it is that presenting them together telegraphs that they are equally weighted and equally deferrable, which they usually are not. Chat interfaces train this habit because chat interfaces expect a paragraph as a natural response. Stack three questions into one message and the answer comes back as a paragraph with three sub-answers threaded through it, and I have to unpick which sentence answered which question, and then check none of them quietly got skipped.
Air traffic control closes the loop differently, and the contrast is instructive. An instruction is not treated as received because it was transmitted — the pilot reads it back, and the controller listens to that readback and challenges it if something is missing or wrong. Nothing that matters is assumed to have landed. That is not because pilots are slow to process a longer message; it is because an item that goes unacknowledged is exactly where things silently drop, and dropping the wrong one is expensive in a way that is not recoverable after the fact. Chat has no equivalent step. A message goes out, something comes back, and nothing checks whether every part of what was asked is actually in it.
The same logic applies to a decision an agent needs from me mid-task. If I get three questions in one turn, I answer the one that is easiest to answer, and the other two sit half-acknowledged until something breaks because the agent proceeded on an assumption for the one I never actually addressed. The cost of a dropped question does not show up at the moment it is dropped. It shows up later, as a decision that was made for me by default rather than by me.
There is a version of this that is worse than three stacked questions, which is one question buried inside a status report. The agent tells me what it did, what it found, and what it is planning to do next, and somewhere in the middle of that sits a genuine decision point that needs my input, phrased like one more sentence of narration rather than a question. I have walked past those more than once, because nothing in the shape of the message told me to stop reading and answer something. A question that does not look like a question does not get treated like one.
Stacking also feels efficient from the sender’s side in a way it is not from the receiver’s side. It looks like one round trip instead of three. What it actually produces is three quick decisions, each made with full attention, replaced by one slow decision made under the load of two others competing for the same attention. The saved round trip is real. The cost of answering worse, or answering the wrong one, is usually bigger than the round trip it saved.
Forcing one question per turn changes what a check-in means — from a status dump I am expected to parse correctly, into a specific fork that has to be resolved before the agent can continue. That specificity is the whole value. A single yes-or-no, or a genuine A-versus-B-versus-C choice, is something I can decide in the time it takes to read it, rather than something I have to hold in working memory against two other open questions while composing one reply that tries to cover all three at once.
The discipline runs the other way too — it forces the agent to sequence its own uncertainty instead of front-loading every open question it can think of into one message because that is less work than working out which one actually blocks progress right now. Most of the time only one of three questions is genuinely load-bearing; the other two could be answered with a sensible default and revisited later if the default turns out to be wrong. A stacked question is often a tell that the agent has not done that filtering, and has handed the filtering job to me instead. The interaction model that actually works for a running agent looks more like clearance-and-acknowledge than it looks like conversation, and that is a discipline the agent has to hold on its own — it is not something you get for free by asking nicely once and hoping it generalises.