# Approvals

As the agent works, it sometimes pauses to check in with you. It might ask a clarifying question when your prompt leaves something open, or request your approval before completing a sensitive action. Both appear right in the Agent Chat, and the agent waits for your answer before continuing so nothing significant happens without your say-so.

## Clarifying Questions

If your request is unclear, the agent asks rather than guesses. A question appears in the chat with a short list of options to choose from, so you can steer the work without breaking your flow.

1. **Read the question and options**: The agent states what it needs to know and offers a few options, each with a short description. For a single-choice question it suggests a sensible default, already selected, so you can move on quickly if it's right.
2. **Pick an answer**: Choose the option you want. Some questions let you select more than one (you'll see "Select all that apply"). If none of the options fit, choose **Other** and type your own answer.
3. **Confirm**: Choose **Confirm** to send your answer. If the agent asked a few questions at once, you'll step through them with **Continue**, then confirm on the last. To skip answering, choose **Dismiss**.

> [!TIP]
> Answering with a clear choice is faster than typing, but **Other** is there when your preferred answer isn't on the list. Use it whenever the options don't capture what you actually want.

## Permission Requests

Glaze apps are **local-first**, and the agent works within your project. When a step would reach **outside** that project, the agent stops and asks before going ahead so nothing happens beyond your project without your say-so.

You'll see a permission request when a step would, for example:

- **Read or write files outside your project**, such as somewhere else on your Mac.
- **Run a command that touches something sensitive** beyond the project folder.

The request shows a short reason and, when there's one, the exact file path or command involved, so you can see precisely what's being asked.

### Allow or Deny

A permission request gives you two choices:

- **Allow** lets the agent take that one action and continue.
- **Deny** stops that action. The agent carries on without it, usually finding another way or telling you it couldn't proceed.

> [!NOTE]
> Each request is a **one-time** decision about that specific action. Allowing one step doesn't hand the agent open-ended access; the next sensitive action asks again.

> [!WARNING]
> Read what's being requested before you allow it. The reason and the path or command are there so you can decide with the full picture, especially when a step would write outside your project or reach the network.

## Approving a Plan

In **Plan mode**, the agent settles the approach with you before writing any code. When it's ready, it presents the plan and asks **"Would you like to approve this plan?"**

- Choose **Yes, approve the plan** and the agent switches into Build mode and carries it out.
- Not quite right? Pick **Other** and type what to change. The agent revises the plan instead of starting on the app, so you can refine the approach as many times as you like before committing.

Approving a plan is your checkpoint on bigger or riskier work: you agree on the shape of the change first, then let the agent build it. For when to plan first and when to build directly, see [Plan & Build Modes](/prompting/plan-and-build-modes).

## FAQ

### Why does the agent ask permission for some things and not others?

Work inside your project happens freely, since that's what you're building. The agent only stops to ask when a step would reach outside the project, such as writing files elsewhere on your Mac or connecting to the network. That keeps your local-first app contained while still letting it do what it needs to when you approve.

### What happens if I deny a permission request?

The agent skips that action and continues. Often it finds another way to reach your goal; sometimes it tells you it can't finish that part without the access it asked for. You can always re-prompt and allow it on the next attempt.

### Can I ignore a question the agent asks?

You can **Dismiss** it, but answering usually gets you a better result, since your answer is exactly the detail the agent was missing. If you dismiss, the agent proceeds with its best guess, which you can correct afterwards. See [When the Agent Gets It Wrong](/prompting/writing-effective-prompts#when-the-agent-gets-it-wrong).

These prompts are part of how you steer the agent as it works. To go further, see [Iterating](/prompting/iterating) for refining over multiple turns, and [Annotate](/prompting/annotate) for pointing at the exact thing you want changed.