Plan Mode
Plan Mode makes the agent show you the design before it touches your dashboard. It reads your data, works out what should go on the canvas, and presents a plan you can approve or send back — nothing is built until you say so.
Use it whenever the request is open-ended ("make me a sales dashboard"). For a one-line tweak, stay in Act mode and let it just do it.
Plan and Act
The mode toggle sits in the message box, and a Planning badge appears while plan mode is on, with the reminder that changes are locked until you approve a plan.
| Mode | What the agent may do |
|---|---|
| Plan | Look — read your sheets, inspect datasets and widgets, check what a chart can be filtered by. It cannot change anything. |
| Act | Build — create datasets, add and arrange widgets, bind data, apply themes. Writes that matter still ask first. |
Switch modes at any point in the conversation. If the agent started building too early, flip back to Plan and ask for a design.
When the plan is ready
The conversation shows a Plan ready · N steps divider followed by the step tree — the actual sequence the agent intends to run, expandable step by step. View full plan opens it larger, together with the agent's written rationale, including a layout sketch when it drew one.
Below the input box, a Plan Ready panel appears:
- Approve & Execute — the agent starts building.
- Revise — send it back to plan. Type what is wrong in the box above the buttons, and the revision follows your feedback rather than a fresh roll of the dice.
The panel stays until you answer. Take your time; nothing executes in the meantime.
What a good plan is telling you
Read it as a checklist before work starts, because this is the cheapest moment to change direction:
- The chart choices — is a stacked bar really what answers your question, or should it be a line?
- The data shape — if the plan creates a derived dataset before charting, that usually means your raw sheet needed reshaping. Good to know either way.
- What is missing — if a filter you expected is not in the plan, say so now.
Reviewer loops
With reviewers active, a finished pass gets checked, and a failing check sends the master back to re-plan the affected part rather than declaring the job done. You see each round in the conversation, so it never quietly churns.
Empty selections are normal
A chart with no segment clicked, or a dropdown with nothing chosen yet, is a valid starting state — the plan does not need to preselect anything. You drive filters from the moment the dashboard is live.
Where to go next
- Agent Perspective Overview — the agent team and how it delegates
- Settings → Permissions — what the agent asks before doing
- Thinking & Reasoning — inspect the reasoning behind a plan