Mid-run questions

Let an agent stop and ask you, instead of guessing on the expensive decisions.

Unattended agents guess, and guesses compound. For steps where a wrong assumption is expensive, let the agent stop and ask:

- type: prompt
  prompt: Clarify the requirements, then implement the task.
  allowQuestions: true

It is per step and off by default. Steps without it are steered to run unattended and never ask.

What asking looks like

When genuinely blocked, the agent asks everything it needs in one go — there is no limit on how many questions, so if you want fewer, say so in the step's prompt. Each question is one of:

• A choice — single or multi-select, with an "Other" free-text escape. Options can be rows, cards with an icon or a picture, or tiles that are pictures: mockups, layouts, screenshots to compare. An option can carry a Recommended-style badge, a colour when it means a cost, a risk or the safe way, and a preview — the code it would write, or what it would look like — shown when you pick it.
• A scale — a slider between two labelled ends.
• A date.
• Your own words — a line or a paragraph.
• Files — data, documents or screenshots only you have.

The set can start with background — text, code or a picture — so you see what the agent sees before you answer. Pictures the agent made are kept with the run, so they're still there when you answer from your phone the next morning, and any of them opens full size — an option's picture from the button on its corner, since pressing the option picks it.

The run parks and the session is held live. Your answer resumes it in place, with its context intact and nothing to restart.

Where you answer

• Browser, desktop, Mac, iPad — a short set is answered right in the run, in the amber Waiting on an answer card. A long one — many questions, big pictures, long code — shows its questions in a line each and Answer in a tab, which opens the whole set at reading width with a Send button that answers the run. The button is on short sets too, if you'd rather have the room. What you pick in one place is there in the other.
• Phone — a push says the agent needs your input; tap it and answer from the run screen, one question per page.
• Slack and Discord — for runs born from a Slack message or a Discord one, the thread gets the question and a link that opens it to answer — unless the run worked with restricted folders or files. It is answered in Peractor, not in the thread.
• Your chat app — if you connected one with Do work or more — not Read only — you can answer from there.

Choose Other to type your own answer or attach files. Each question's Other has its own Attach button and file chips. A file can be the whole answer; it is sent with that question when you submit.

Answering a run you started takes Start runs; answering somebody else's, or one a trigger or an API key started, takes Manage runs. A run that worked with restricted folders or files is answered only by people who can open them.

Answers you give are always carried forward into the run's later steps, whether or not the step remembers anything else.