Skip to main content

Agent Creation Approvals

When the approval workflow is enabled in Settings, agent creation requests are queued for human review instead of being executed immediately. The Approvals page lets administrators examine proposed configurations, assess risk, and approve or deny each request.

Accessing Approvals

Navigate to System Admin > Agent Self-Service > Approvals.

A badge next to the page title shows the number of pending requests when there are items awaiting review.

Filtering Approvals

Use the status filter buttons at the top of the page to view requests by status:

FilterDescription
PendingRequests awaiting administrator review (default view)
ApprovedPreviously approved requests
DeniedPreviously denied requests
AllEvery request regardless of status

Approval Queue Table

The approvals table displays the following information for each request:

ColumnDescription
Proposed AgentThe name and system identifier of the agent to be created. An SME badge appears if the request is for a Specialist Agent.
Requesting AgentThe existing agent that initiated the creation request
Risk ScoreAI-generated danger assessment (see below)
StatusCurrent status: Pending, Approved, or Denied
RequestedDate and time the request was submitted
ActionsClick Review to open the detail modal

Reviewing a Request

Click Review on any request (or click the table row) to open the review modal. The modal displays comprehensive information to help you make an informed decision.

Proposed Agent Details

  • Name - The system identifier for the new agent (e.g., marketing_content_agent)
  • Display Name - Human-readable name shown in the admin console
  • Type - Whether this is a standard agent or a Subject Matter Expert (SME)
  • Requested By - Which existing agent initiated the creation
  • Risk Score - Color-coded danger assessment

Proposed Instructions

The full system prompt that the new agent would use. Review this carefully for:

  • Appropriate scope and boundaries
  • No references to sensitive data access beyond what's needed
  • Clear, well-defined responsibilities
  • No instructions that could lead to harmful behavior

Requested Tools

The list of tools the new agent would receive. Remember that an agent can only grant tools it already has. Verify that the requested tools are appropriate for the agent's stated purpose.

AI Safety Analysis

The system provides AI-generated analysis to assist your review:

  • Analysis Summary - A brief assessment of the overall risk level and reasoning
  • Recommendations - Specific suggestions for improving safety (if any)
  • Flags - Highlighted concerns that warrant attention (shown in red)

Risk Score Indicators

The risk score uses color-coding for quick assessment:

ScoreColorMeaning
Below 30%GreenLow risk - instructions appear safe
30-69%YellowModerate risk - review instructions carefully
70% and aboveRedHigh risk - exercise caution, may warrant denial

Making a Decision

Approving a Request

  1. Review all the information in the modal
  2. Optionally add review notes explaining your decision
  3. Click Approve
  4. The agent is created according to the proposed configuration
  5. If "Create Agents as Inactive by Default" is enabled in Settings, the new agent will need to be manually activated

Denying a Request

  1. Review all the information in the modal
  2. Add review notes explaining the reason for denial (required for denials)
  3. Click Deny
  4. The request is marked as denied and logged
info

Denial notes are recorded in the audit log and may be visible to the requesting agent in future executions, helping it understand why the request was rejected and adjust its approach.

Viewing Previous Decisions

For approved or denied requests, the review modal shows the decision details:

  • Reviewed By - The administrator who made the decision
  • Reviewed At - When the decision was made
  • Notes - The reviewer's notes or reason for denial

Tips

  • Check approvals regularly - Pending requests block agent workflows, so review them promptly
  • Use the risk score as a guide - Low-risk requests from trusted agents can be approved quickly
  • Read flagged instructions carefully - If the AI safety analysis raises flags, take extra time to review
  • Add meaningful notes - Your notes create a valuable audit trail and help agents learn from rejections
  • When in doubt, deny - It's safer to deny and investigate than to approve a questionable request