Skip to main content
A 3-way selector controls how the agent handles permission for the rest of a task.
Changed in v3. It now lives in a tray that slides open when you tap the + next to the chat input, alongside the image-attach button and, in v3.3, the inline effort slider. The plus springs into × while open. It is no longer a bar permanently pinned under the composer.
Normal, Plan, and Auto mode selection above the v3.3 chat input, alongside the reasoning effort slider

Normal

The default. Every mutating tool call asks first, exactly as described in Tools & Permissions.

Plan

The agent proposes a plan — a short, numbered list of the tool calls it intends to make and why — before touching anything, and waits for you to Approve & Run or Cancel.

Auto

Skips every permission prompt for the rest of the task, including execute_luau, which normally always asks no matter what.

​
Plan, in detail

With Plan selected, the first reply to a new task is always a plan instead of a tool call — the model is explicitly instructed not to act yet. The plan renders in a blue card with two buttons:
  • Approve & Run — re-sends the conversation with the plan gate lifted for that one continuation, so the agent now actually executes the calls it described.
  • Cancel — leaves the plan sitting in the transcript; nothing further happens.
If a task genuinely needs no tools at all, the agent just answers normally — Plan only gates turns that would otherwise call something.

​
Auto, in detail

Auto Mode is the one mode that requires a deliberate, explicit confirmation to turn on. Selecting it opens a dialog spelling out exactly what it bypasses — including execute_luau, the one tool that’s permission-locked in every other mode. Cancelling leaves you on whichever mode you were on before.
Once confirmed, every tool call — read or mutating — runs immediately with no prompt, for as long as Auto stays selected. An explicit Always Deny you’ve set on a tool is still honored; Auto Mode skips the ask flow, not a deny you configured on purpose. Auto Mode also changes what a spawned subagent is allowed to do: if the main session is in Auto when spawn_subagent is called, the subagent inherits full unattended access, execute_luau included.

​
Switching mid-task

The selector applies to new tasks — a task already in flight keeps the mode it started under. Switching from Plan to Normal (or vice versa) doesn’t need confirmation; only switching into Auto does.