> ## 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.

# Automatic review and escalations

> Understand how SuperEmber reviews exact actions, denies unsafe proposals, and pauses for human decisions.

SuperEmber reviews boundary-crossing tool calls before execution. The goal is to preserve useful autonomy without treating every agent instruction as unlimited authorization.

## What is reviewed

Review applies to actions across boundaries such as:

* connected-app writes
* MCP calls
* schedules and persistent automation
* sandbox commands and processes
* public sharing
* skill installation or removal
* agent-to-agent messaging and delegation
* external network access
* knowledge writes

The exact surface depends on the agent's enabled capabilities and workspace policy.

## Decision lifecycle

```mermaid theme={null}
flowchart TD
    P[Agent proposes an exact tool call] --> B[Bind call to tenant, turn, tool, and action digest]
    B --> R{Automatic review}
    R -->|Approve| E[Execute exact call]
    R -->|Deny| S[Do not execute; choose safer route]
    R -->|Escalate| H[Park exact call for a human]
    H -->|Approve once| E
    H -->|Deny| S
    E --> A[Record execution result]
```

The reviewer considers user intent, authorization, effect, and risk. Sensitive arguments are redacted and bounded before review or audit storage.

## What each decision means

| Decision               | Result                                                                                                   |
| ---------------------- | -------------------------------------------------------------------------------------------------------- |
| Approved automatically | The exact reviewed call may execute.                                                                     |
| Denied automatically   | Nothing executes. The agent must choose a materially safer route rather than rephrasing the same action. |
| Human required         | The exact call remains parked until a person approves or denies it.                                      |

Repeated denials in one turn trigger a human decision instead of allowing an agent to keep proposing near-identical actions.

If automatic review is unavailable or cannot preserve the review record, SuperEmber fails closed to human approval.

## Delegated work has stricter authorization

A delegation is authored by another agent, not directly by a human. It does not grant explicit authorization for destructive changes, permission changes, financial or legal commitments, credential access, or other high-impact actions.

When a delegated agent needs a human decision:

1. The target team's Main operation thread changes to **Waiting for a human decision**.
2. A review-needed receipt appears in that operation thread.
3. The source conversation receives a waiting receipt and a link when available.
4. After approval or denial, both receipts update in place.
5. Approval resumes the parked call; denial tells the agent to choose a safer route or report the blocker.

If the target has no Main channel, the action remains paused and the source receives an explicit notification failure message.

## Approval is exact and one-use

Before approving, verify:

* the tool and operation
* the account, recipient, or target system
* the arguments and scope
* the expected external effect
* that the action still matches the user's request

Approval applies to the parked call only. Changing the call after review invalidates the decision, and an executed call cannot be replayed through the same approval record.

## Automatic review is not the workflow inbox

Automatic review governs dynamic agent tool calls. Workflow approvals are designed steps in a durable process. Personal Agent connector approvals appear in **My agent**, while delegated human escalations appear in team operation threads.

See [Approvals, review, and inbox](/product/approvals-and-inbox) for the full surface map.