Skip to main content
A plan card is the agent’s proposal for multi-step work. It appears in the chat as a titled, numbered list of steps. Each step can carry a model chip, a short reason, and an approval checkpoint. A Why this plan disclosure explains the overall approach. Simple requests do not produce a plan card. The agent runs them directly.

​
Read the card

​
Edit the plan

You can edit the card directly before you run it, and you can edit the remaining steps while it runs.
  • Rename a step by clicking its label.
  • Reorder steps by dragging them.
  • Add or remove steps with the controls on each row. Removing a step also removes the steps that depend on it.
  • Pin a model on a step by clicking its model chip. A pinned model is final. The agent does not replace it.
  • Toggle an approval checkpoint on any step.
The agent sees your edits. When it continues, it works from the edited plan, not its original proposal.

​
Run, cancel, or revise

  • Run plan starts execution. Every remaining step becomes a visible row in the queue, so you can see the whole tail.
  • Cancel plan tells the agent you do not want this plan. The agent treats it as feedback and proposes a revision.
  • Not happy with the plan? Type a change in the text box on the card. This sends a normal message and produces a new plan version.
Keyboard: Cmd+Enter (Ctrl+Enter) submits the revision box.

​
Approval checkpoints

A checkpoint is a hard stop. The agent cannot run past an unapproved checkpoint, even if it wants to. When execution reaches one, the chat row in the sidebar shows Needs approval, and the queue row waits. Approve it on the card to continue. Approval is consumed once, so the same step does not ask twice. Checkpoints are separate from spending caps, which gate on cost.

​
Plan versions

One chat has one logical plan with a version history.
  • A re-render in the same turn updates the card in place.
  • A re-plan in a later turn creates a new card, v2, v3, and so on, at the bottom of the chat.
  • The old card freezes as superseded and shows Jump to the latest plan version.

​
Steps that continue automatically

When a step’s run lands, the agent continues with the next step without a message from you. The card shows the next step as Up next. If a batch lands partially, the agent continues with the runs that succeeded and reports the rest. You can interrupt at any time. See Queue and steering for how a message sent mid-run is handled, and how to halt or resume a chain.