Structured Questions

Desktop app
All plans
Checked Oct 22, 2026

What it does

Agents stop asking their questions in the middle of their answer. When an agent needs a decision, a preference or a missing fact, it shows a card docked right above its input box: one question at a time, each with clickable options and a one-line description of what every option implies, and, when the agent compares designs, a mockup of each option (plain text, or HTML drawn in an isolated frame). The terminal stays visible and usable above it. The agent waits for the answer instead of carrying on with a guess. It works the same with every agent CLI, because the question goes through AgentsRoom rather than through the CLI's own dialog.

Video (1 min 13 s): Structured questions, shown in the real app. A Codex agent is asked to add a newsletter signup form and to ask first where the emails should be stored: the card A question for you opens above its input box with three explained options, the agent shows as waiting (1 needs input, Needs you at 1), one click on Shop database then Skip and send answers, and the agent opens its reply with "Where should the newsletter signup email addresses be stored? → Shop database" before carrying on. A later question shows Paused in the card header while the card stays open.

Where to find it

It is available to everyone and on by default for every agent. Turn it off in Settings > Agent behavior > Structured questions.
For one project only: Project settings > Default agents > Structured questions (with Reset to follow the global switch again).
For one agent only: Edit agent > Advanced options > Structured questions: Same as Settings, On: asks through cards or Off: asks in the conversation.
Or straight from a card: its ... menu, Stop asking with cards, then For this agent, For this project or For all projects. A message says where to turn cards back on, with Undo.
The card appears above the input box of the agent that asks, titled A question for you, with the agent's name, the step number and Pause in 0:45 counting down in its header. While it waits, the red 1 needs input pill shows at the top of the window and the Needs you counter of the sidebar's Running agents goes up. You are told the usual way an agent needs you: the red dot on its tab and in the sidebar, the Needs input inbox, the dock badge and a push on your phone.
Without opening the agent: the Needs input list at the top of the window shows the question under the agent's name, with its options when one click answers it.
In the Open space: the waiting agent's card shows the question in the Needs you column, and selecting it (or Reply) docks the whole card above the Open space's input box, answered exactly as in the agent's console.
On the phone: the same card above the agent's input box in the AgentsRoom app, and answer buttons on the lock screen (iPhone Live Activity, Android live notification) when the question allows it.
Through the personal assistant (see Assistant (personal assistant)): ask it "what are my agents asking me?" from the phone, by voice or in its panel. It reads every open card of every project (agent, project, question, options, the agent's recommendation), then passes on the answer you give it ("answer yes to the security architect"), or a Not clear? when you say the question is unclear. It never picks for you, not even the recommended option, and answers one question at a time. Since 2026-10-20.
In an agent team run: every agent of the run asks with the same card during its step, on its own console. The run waits on that step (after a few minutes its panel says the step is waiting for you), and its teammates' messages wait until you answer. Where the run goes after a step is still decided in the run's panel (Wait for me step, see Agent Teams).

How to use it

Open the agent that is waiting. Click an option: for a single-choice question, the click is the answer and the next question appears.
Or type in the card's own field, under the options: your own words answer the question on screen, and option numbers pick options ("2", or "1, 3" for several). Press Enter to send. The input box below the card is NOT part of the card: what you send there reaches the agent as a normal message, and the question stays open.
Several choices allowed, or mockups to compare: tick the options, then Next. With mockups, hover an option (or move the keyboard focus onto it) to see its mockup; the click only chooses. The expand button in the mockup's header (Enlarge preview) opens it in a large window that scrolls both ways, to read a long draft or a full page; Esc, the cross or a click outside closes it. Only the question and its options scroll: the card's field and Next stay visible below them, and a fade at the bottom of the list shows that more options are hidden.
On the last question left, answering does not send the card yet: two buttons appear, Send (sends your answers now) and Add a note (opens Additional notes, a last field for a remark no option covered; Enter or Send sends everything with it, Back returns to the question). The numbered steps in the card's header go back to an earlier question (the notebook step opens the notes once every question is answered). Words typed in the card's field are enough, with no option ticked: Enter sends the card at once, and Send / Add a note become active as soon as the field holds text.
The agent may mark the option it advises with a Recommended badge. It is only advice: pick any option.
Annotate (on the question's line) opens a small note field under the options. Your note is sent with your answer to THIS question, for example "yes, but keep the old version for a month". It does not answer on its own: you still click an option (a click on a single-choice option sends the note along). The note stays if you fold the card or go to another question; the cross removes it.
Not clear? (on the question's line) replaces the card's field with "What isn't clear?": say what blocks you if you like (optional), then Ask for an explanation (or Enter; Esc goes back). The card closes, the agent explains the question in plain words, then asks it again, reworded, with its advice marked. The answers you already gave to the other questions are kept and not asked again.
Several answers (on the question's line, only where the agent asked for one answer): the options become checkboxes, tick several, then Next (or Send on the last question). Click it again to go back to one answer.
Each option also has its own two icons, shown when the pointer is over it: Annotate this option (a remark on that option, sent with your answer whether you pick it or not) and This option isn't clear (the agent explains that option, then asks again). On the phone, long-press an option to get the same two actions.
The arrow at the right of the header folds the card to one line; the question stays open. The cross next to it closes the card without answering (the same as Close without answering in the ... menu), folded or not.
The ... menu: Answer in the conversation instead closes the card (the agent is told and asks again in the conversation if it still needs to), Close without answering closes it with no answer at all (the agent does not ask again, goes on with its best judgment and says what it chose, or ends its turn), and Stop asking with cards turns the cards off (see above): the agent then asks again in the conversation.
From the Needs input list (the red "needs input" pill at the top of the window, or Cmd+Shift+I to jump): a single question with single choice and no mockups is answered by clicking one of its options right there. Anything else shows the question and Open the agent. Not clear? works there too, in one click.
From the phone: open the agent in the AgentsRoom app. The card sits above its input box: tap an option, or type your answer (or option numbers) in the card's own field, exactly like on the computer. The app's input box still sends a normal message to the agent. The numbered steps go back to an earlier question. The arrow (or a tap on the card's top line) folds the card to one line, and the cross in its top corner closes it without answering, folded or not; it never takes more than part of the screen, and its bottom line always shows Answer in the conversation instead (the agent asks again in the conversation) and Close without answering (the agent goes on without an answer). HTML mockups are not drawn on the phone: the card says to view them on the computer. Recommended, Annotate and Not clear? are there too (with a desktop app recent enough to read them).
From the lock screen: when the card holds a single question with two options short enough for a button (the agent can give each option a short button text), the two buttons answer it (Face ID on iPhone, unlocked phone on Android). Otherwise the banner shows the question with Open, which opens the card in the app.
From the Open space (Cmd+Shift+O): the card of an agent waiting on a question shows the question itself. Click the card, or Reply, and the same card docks above the Open space's input box, with its steps, options, own field, Annotate, Not clear?, ... menu and cross: everything works as in the agent's console, including Send / Add a note on the last question. The Open space's input box still sends a normal message to the agent.

Whichever screen answers first wins: the card disappears everywhere else. An answer given on a screen that was not up to date (the card had moved on in the meantime) is refused rather than applied to another question.

After you answer, the agent starts its reply by restating each question and your answer, so the exchange stays readable in the terminal.

The agent only waits a short time for the card: the header shows Pause in 0:17 (hover it for the details). At zero the card does NOT close: the agent pauses, the header says Paused, and the agent stays marked as waiting for you (red dot, Needs input). Past that it pauses, the card stays, and when you answer later your answer is typed to the agent as a message starting with "[AgentsRoom] Answer to your question". Nothing is lost if you are away.

Settings

askUserEnabled (global, Settings > Agent behavior > Structured questions): agents ask through cards. Default on. Applies at each agent's next launch. Followed on every computer of your account.
askUserEnabled (project, Project settings > Default agents): overrides the global switch for one project; empty follows it. Followed on every computer of your account.
askUserEnabled (agent, Edit agent > Advanced options): pins the value for one agent; empty follows the project, then the global switch.
The card's ... menu writes these same three settings.
askUserWaitSeconds (global, same section, Wait before the agent pauses (seconds)): how long the agent waits for the card before pausing, 10 to 600, default 45. Codex and the CLIs whose own limit is unknown stop waiting after about 50 seconds whatever the value: their tool calls are cut at 60 seconds. oh-my-pi stops after about 20 seconds: it cuts a tool call at 30 seconds.
askUserHtmlPreviews (global, Draw HTML previews): draw the HTML mockups agents attach to their options. Off shows their source code. Default on.

Agent tools (MCP)

ask_user: asks the user one to four questions as a card docked on the caller's input box and waits. Listed only to the agents it is on for (checked when the agent starts). Never listed on Aider, which has no MCP client. Each option can carry a short text (22 characters at most) used as the lock-screen button when its label is longer.
settings_set (scope: "global", "project" or "agent"): an agent can turn the feature on or off when you ask it to.
agents_questions_list and agents_question_answer: the personal assistant reads the open cards of every project and hands over the answer you gave it. Listed to the assistant only (hidden from every other agent, refused to its specialists); it never sees nor answers its own cards.

Providers

All providers, through the AgentsRoom MCP server. On Claude Code, an agent that asks with cards no longer loads Claude's own question dialog (one tool less in every session, so less context used); it keeps it when cards are off for it.

The difference is how long a CLI lets one tool call wait: Claude and Grok wait as long as the setting says; Codex and the other CLIs stop at about 50 seconds (oh-my-pi at about 20), after which the agent pauses and receives your answer as a message.

Mobile

The card is on the phone too, above the agent's input box in the AgentsRoom app: same title, same steps, same Next / Send, same Answer in the conversation instead and Close without answering. No Additional notes step there: the phone, the lock screen and the Needs input list send the answers right away. The card has its own answer field like on the computer; the input box stays a normal message to the agent. It folds to one line and scrolls inside a part of the screen, so a long question never hides the input box. HTML mockups are only drawn on the computer; a text mockup has the same Enlarge preview button, which opens it full screen. The Recommended badge, Annotate and Not clear? work on the phone card like on the computer (the buttons only appear when the desktop app is recent enough). On the lock screen (iPhone Live Activity, Android live notification), a single two-option question gets its two buttons, whatever CLI the agent runs; any other card shows Open. Needs the phone app version that ships this card.

Limits

Up to 4 questions per card and 8 options per question.
A change of setting applies at the agent's next launch (its tool list and instructions are set when it starts).
HTML mockups run in an isolated frame with no access to the app or your files; scripts inside them may still load content from the internet.
Agents started by a webhook trigger (restricted runs) never ask with cards: nobody is there to answer.
An agent running on another of your computers (see Remote Fleet) shows its cards too since 2026-10-20, on the desktop that watches it (Terminal and Conversation views) and on the phone. The answer is applied by the computer that runs the agent; the "stop asking with cards" settings stay on that computer and cannot be changed from here.

Common questions

How do I make my agent ask me before it guesses? Nothing to set up: with Structured questions on (the default), an agent that hits a real choice asks it in a card above its input box and waits. You can also say so in your task ("Before writing any code, ask me where the emails should be stored"). Click an option, then Send (or Add a note first). Video: https://youtu.be/UD6U--rX1w8
Codex asks a question and keeps going without waiting for me. Can I make it stop and wait? Yes, with Structured questions (on by default): the agent asks through a card and pauses until you answer.
Can I ask my assistant what my agents are waiting on, and answer by voice? Yes. Ask the personal assistant "what are my agents asking me?": it reads you each open question with its options and the agent's advice. Tell it your choice ("answer the second option to the backend agent") and it passes it on. It only sends what you said; when it is unsure which card or option you mean, it asks you first.
Can I answer an agent's question without the arrow keys of the terminal? Yes, with Structured questions on, every option is a button in the card.
I did not answer in time, is my answer lost? No. The agent pauses and your answer is sent to it as a message as soon as you give it.
Can I turn it off, or keep it for only one agent? Yes: turn it off in Settings > Agent behavior > Structured questions, and set the agent you want to keep it on to On: asks through cards in Edit agent > Advanced options > Structured questions. The same field set to Off: asks in the conversation turns it off for one agent only.
Can I stop the cards for one project only? Yes: from a card, ... > Stop asking with cards > For this project, or in Project settings > Default agents. For this agent and For all projects are in the same menu.
Does the card block the terminal? No. It sits above the input box and pushes the terminal up; fold it to one line with the arrow in its header.
Can I add a remark that none of the options covers? Yes: on the last question, click Add a note instead of Send, type it and press Enter. The agent reads the note next to your answers. For a remark on one specific question, use Annotate on that question.
What does the Recommended badge mean? The agent advises that option. It is a suggestion, nothing is chosen for you.
Can I add a remark to one question without typing my whole answer? Yes: Annotate on the question's line. Write the note, then click your option as usual: the note leaves with that answer. On the phone it works the same way.
The agent asked for one answer but several apply. Can I pick more than one? Yes: click Several answers on the question's line, tick the options, then Next or Send. Same button on the phone.
Can I comment on, or ask about, one option only? Yes: hover the option and use its icons (note, question mark); on the phone, long-press the option.
I don't understand the question. Can the agent explain it? Yes: Not clear? on the question's line, optionally say what blocks you, then Ask for an explanation. The agent explains, then asks the question again in plainer words. Your answers to the other questions are kept.
How do I compare the mockups without choosing one? Hover each option: its mockup shows next to the list. Clicking an option chooses it (or ticks it, when several choices are allowed). On the phone there is no hover: a tap chooses.
Can I answer by typing? Yes, in the card's own field: your own words, or option numbers like "2" or "1, 3", then Enter.
Can I send the agent a message while a question card is open? Yes. The input box below the card sends a normal message to the agent, right away, on the computer and on the phone. The question is not answered by it: the card stays open until you answer it on the card.
Can I answer an agent's question without opening the agent? Yes: the Needs input list at the top of the window answers a single-choice question in one click (for several questions, several choices or mockups, it offers Open the agent), and the Open space shows the whole card above its input box for the agent you select there.
Can I answer from my phone? Yes: the card shows above the agent's input box in the AgentsRoom app, and a two-option question can be answered from the lock screen.
Why does my lock screen only show Open? The lock screen fits two buttons and no typing: it answers a single question with two options whose text fits a button. Everything else opens the card in the app.
I don't want to answer the question. What do I do? Nothing forces you to: fold the card (arrow in its header), or close it with the cross at the right of its header, or from the ... menu on the computer (at the bottom of the card on the phone). Answer in the conversation instead tells the agent you prefer to talk in the conversation; Close without answering tells it no answer is coming, so it does not ask again and carries on with its own choice (or stops its turn). Ignoring it is fine too: the agent simply keeps waiting, marked as needing input.
I typed in my phone's input box while a question card was open. Where did my message go? To the agent, as a normal message, like on the computer. The card stays open: answer it with its options or its own field.
I answered the agent's question in the conversation. Why is the card still open? A message typed in the input box is a normal message to the agent, on the computer and on the phone: it does not answer the card. Once the question is settled, close the card with the cross at the right of its header (Close without answering): it disappears from every screen, the computer, the phone and the lock screen.
I answered on the computer and on the phone at the same time. Which answer counts? The first one. The other screen's card disappears, and an answer sent for a question that already moved on is refused.
An agent of my team run asked me something in its terminal and nothing alerted me. Fixed on 2026-10-19: agents of a team run could not use the card, so they asked in plain text and showed as idle. They now ask with the card during their step: red dot, Needs input list, push and phone card, like any other agent. With the cards turned off, a team agent asks through the run instead, at the end of its step (a Wait for me step or the needsInput flag).
Is there a video that shows it? Yes, a one-minute demo in the real app: an agent asks where to store newsletter emails in a card, one click answers, and the agent restates the answer and carries on: https://youtu.be/UD6U--rX1w8

Conversation View: shows the questions Claude asks with its own dialog as cards too.
Command Proposal: the same kind of card when an agent asks you to run a command.
AgentsRoom MCP: the MCP tools agents use to drive AgentsRoom.