> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mobilerun.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Human in the loop

> When your agent asks you a question or waits for your approval before doing something sensitive, and what happens while you decide.

Your agent does not guess on things that matter, and it does not act on sensitive requests without checking with you first. Both show up as a card in the chat.

## When your agent asks a question

Your agent asks when it needs information only you can give, or when a step is genuinely ambiguous, such as which of two matching accounts you meant.

A question can arrive as option cards you tap to answer, free text you type yourself, or both together. Answer directly in the chat. There is nothing else to configure.

> Which of these two events did you mean: "Design Meetup" on the 12th, or "Design Meetup Berlin" on the 19th?

Your agent does not ask for confirmation on everything, only where the answer actually changes what it does next.

## What needs your explicit approval

Some actions are sensitive enough that your agent always stops and shows you an approval card before doing them, no matter how the request was phrased:

* Creating a Cloud Android subscription phone
* Resetting a device to a clean state
* Stopping or resuming a pay-as-you-go phone
* Permanently removing a temporary, pay-as-you-go phone
* Cancelling a managed proxy or deleting one you brought yourself
* Deleting a recording
* Deploying a routine, and aborting, unblocking, enabling, or archiving one
* Turning off a routine's health monitoring or its notifications
* Creating or archiving an agent, or changing an agent's instructions
* Every write to a connected service, such as sending an email or editing a file

Creating a pay-as-you-go phone is different: when you explicitly ask for one, that request counts as your consent and your agent creates it. It asks you to confirm per-minute billing in the conversation only when your request doesn't clearly accept it.

### Auto mode

The chat composer has a permission mode. In **Auto** mode your agent runs the listed actions without asking. Some cards always ask, whatever the mode: creating or archiving an agent, changing an agent's instructions, archiving or enabling a routine, and turning a routine's health monitoring off. Keep the default mode if you want every sensitive action to stop for you.

> [!WARNING]
>
> These are one-way or billable actions. Your agent tells you what it is about to do in one short sentence before the approval card appears, so it never shows up out of nowhere.

You can approve or decline. Approval applies only to that one request; the next reset, shutdown, or deletion asks again. Declining stops the action; nothing happens. Everyday, reversible steps, like installing an app or filling out a form, do not need this card. It is reserved for actions you cannot casually undo.

## Saving a login

If a task needs you to log into an app for the first time, your agent does not ask you to type your password into the chat. It shows a secure card where you enter your credentials directly. Afterward your agent only knows a reference to them, never the actual value.

> Save my login for this app so future runs don't need to ask me again.

If a session on the device is already active, your agent skips the login step entirely rather than asking you to save anything.

## If you do not answer

A question waits for you, and the job is not marked as failed while it is paused. An approval card expires after about ten minutes; if you come back later, send a new message and your agent asks again. A single turn runs for at most two hours.

## Leaving the tab and coming back

If the job is running on a phone, closing the chat does not stop it. Work already in progress on the device continues.

When you reopen the conversation, it catches you up automatically, including any question or approval that is still waiting for you. You do not need to scroll back through everything that happened while you were away.

If you would rather not check back manually, set up notifications so you hear the moment your agent finishes, fails, or needs your answer to a question. See [Notifications](https://docs.mobilerun.ai/agents/notifications.md).

## Stopping a running job

Ask your agent to stop, or use the stop control in the chat. This ends the current turn. Anything your agent already did on the phone stays done; it just does not take any further action.

## Blockers on the phone

While working on a device, your agent sometimes runs into something covering the screen that it did not ask for:

| What it looks like | What your agent does |
| - | - |
| An "app isn't responding" box, a rating nag, or an update prompt | Dismissed automatically, then it continues. |
| A permission prompt clearly tied to what you asked for (like camera access for a scan) | Granted automatically. |
| A permission prompt that is not obviously needed | Surfaced to you as a question before granting it. |
| A request for contacts, SMS, call log, or location access | Always surfaced to you as a question, regardless of the task. Your agent never grants these on its own. |
| An unrecognized popup with no clear close button | Your agent stops and tells you, rather than guessing or backing out blindly. |

Your agent never scrolls past or blindly dismisses something it does not recognize. If a login screen appears unexpectedly mid-task, it stops rather than retrying, since the next tap could go somewhere wrong.

## Related

> [card] [Notifications](https://docs.mobilerun.ai/agents/notifications.md)
>
> Get notified when your agent finishes, fails, or has a question for you.

> [card] [Devices](https://docs.mobilerun.ai/agents/devices.md)
>
> Phones, accounts, and credentials your agent works with.

> [card] [Routines](https://docs.mobilerun.ai/agents/workflows.md)
>
> How approvals and questions fit into recurring automation.

> [card] [Limits and costs](https://docs.mobilerun.ai/agents/limits-and-costs.md)
>
> What your agent will refuse or hand off entirely.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.