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

# Approvals, review, and inbox

> Know where SuperEmber pauses work and how to make an informed decision.

SuperEmber has several review surfaces because not every pause means the same thing. Always decide from the surface that owns the work.

<img src="https://mintcdn.com/superember/63eEyx3R1ZkIzyFQ/images/onboarding/workflow-inbox.png?fit=max&auto=format&n=63eEyx3R1ZkIzyFQ&q=85&s=b5a90da6de911c19c06e7405c3a399cc" alt="Workflow inbox for approval, input, choice, and channel-reply waits" width="1280" height="720" data-path="images/onboarding/workflow-inbox.png" />

## Review surface map

| Surface                  | What caused it                                            | Where you decide                                                 | What resumes                     |
| ------------------------ | --------------------------------------------------------- | ---------------------------------------------------------------- | -------------------------------- |
| Automatic tool review    | An agent proposed a boundary-crossing action.             | Usually no interruption; unsafe actions are denied or escalated. | The exact tool call.             |
| Delegated escalation     | A delegated specialist needs a human decision.            | Target team's Main operation thread.                             | The parked specialist call.      |
| Personal Agent approval  | Your agent prepared a consequential connected-app action. | **My agent → Activity**.                                         | One stored connector action.     |
| Workflow approval        | A workflow reached a configured human step.               | **Workflows → Inbox**.                                           | The same versioned workflow run. |
| Channel binding approval | A channel is being added or replaced.                     | Connect flow or administrator approval path.                     | The requested channel binding.   |

## Automatic review

Automatic review can approve, deny, or escalate an exact action. A denial causes no external change. An escalation parks the call for a human instead of guessing.

See [Automatic review and escalations](/product/auto-review-and-escalations).

## Personal Agent approvals

Personal Agent approvals are owner-scoped. Other members and Team Agents cannot approve actions for your private agent.

To decide:

1. Open **My agent**.
2. Expand the approval's review details.
3. Check the connector, action, target account, and arguments.
4. Choose **Approve** to allow that action once, or **Reject**.

Pending approvals expire. Repeated clicks cannot execute the same stored approval twice.

## Delegated escalations

Delegated work uses the target team's Main operation thread so the responsible team sees the action and its context. The original conversation receives a waiting receipt and later updates to the final decision.

Approve only the displayed action. The source agent's request is context, not blanket authorization for the delegated agent.

## Workflow inbox

The workflow inbox collects durable waits that require a person:

| Item           | Decision                                                  |
| -------------- | --------------------------------------------------------- |
| Human approval | Approve or reject an exact workflow step.                 |
| Human input    | Provide the missing information.                          |
| Human choice   | Select one configured option.                             |
| Channel reply  | Answer the question from the linked conversation surface. |

To resolve an item:

1. Open **Workflows → Inbox**.
2. Review the workflow, run, node, request, and expiration.
3. Make the requested decision or provide the input.
4. Open the linked run and confirm that the same run resumed.

## A good approval question

An approval should answer one question clearly: **Should this exact step continue?**

Require a human decision when the action is externally visible, destructive, permission-changing, financially or legally meaningful, credential-sensitive, or outside clear user intent.

Do not add approvals to every harmless read. Excessive prompts train people to approve without reviewing.