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 up to four multiple-choice questions — each with a short label, two to four options with descriptions, single or multi-select, plus an "Other" free-text escape.

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

Editor — the question panel docks under the live session's action bar, one question at a time.
Browser, desktop, Mac — the run page shows all the questions at once.
Phone — a push says the agent needs your input; tap it and answer from the run screen.
Slack — for runs born from a Slack message, a single-choice question posts native buttons right into the thread.
Your chat app — if you connected one with the Act scope, you can answer from there.

A run waiting on answers keeps its workspace alive, so it counts against your concurrency cap — unlike a run parked at a gate. Answer or cancel it rather than leaving it overnight.

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