Skip to main content

Agent Self-Service Overview

Allow your AI agents to autonomously create new agents and Specialist Agents (SMEs) based on operational needs, with configurable safety controls, approval workflows, and audit logging.

What is Agent Self-Service?​

Agent Self-Service enables orchestrator-style workflows where an existing AI agent can programmatically spin up new agents to handle specialized tasks. For example, the AI Chief of Staff (AICOS) can create SME agents on demand when it identifies a gap in your workforce's coverage.

This feature is disabled by default and must be explicitly enabled by an administrator.

Key Capabilities​

  • Autonomous Agent Provisioning - Agents create new agents or SMEs without manual setup
  • Approval Workflows - Optionally require human review before agents go live
  • AI-Powered Safety Analysis - Every creation request is scored for risk using AI instruction analysis
  • Rate Limiting - Configurable quotas prevent runaway agent creation
  • Templates - Pre-built blueprints for consistent, quick provisioning
  • Complete Audit Trail - Every creation attempt is logged with full parameters and outcomes

Benefits​

BenefitDescription
Dynamic ScalingAgents create specialized helpers as workload demands change
Operational AgilityNew capabilities can be deployed without admin intervention
Safety FirstAI-powered risk scoring and optional human approval gates
Full VisibilityDetailed logs of every creation attempt, approval, and denial
Template ConsistencyPre-configured templates ensure new agents follow org standards

How It Works​

When an agent uses the create_agent or create_sme tool, the following process occurs:

  1. Validation - Parameters are checked for correctness (name format, required fields)
  2. Feature Check - The system verifies Agent Self-Service is enabled for the tenant
  3. Quota Check - Rate limits (per hour, per day, total) are verified
  4. Tool Inheritance - Requested tools are validated against the creating agent's own toolset (an agent can only grant tools it already has)
  5. Safety Analysis - AI analyzes the proposed instructions and generates a danger score
  6. Approval Gate - If approval is required and the danger score is above threshold, the request enters the approval queue
  7. Agent Creation - The new agent or SME is created (as inactive if configured)
  8. Logging - The entire process is recorded in the audit log
info

Agents can only create agents within their own tenant. Cross-tenant agent creation is not supported.

Prerequisites​

Before using Agent Self-Service, ensure the following:

  1. Admin Access - You must have System Admin permissions to configure settings
  2. Active Agents - At least one active agent that will use the creation tool
  3. Tools Assigned - The creating agent must have the create_agent or create_sme tool assigned to it

All Agent Self-Service pages are located under System Admin > Agent Self-Service in the left sidebar:

PagePurpose
SettingsEnable the feature, configure approval workflow, rate limits, and safety threshold
ApprovalsReview pending creation requests, approve or deny with notes
TemplatesCreate and manage reusable blueprints for agent provisioning
LogsView the complete audit trail of all creation activities

Quick Start​

  1. Navigate to System Admin > Agent Self-Service > Settings
  2. Toggle Enable Agent Creation Tool to on
  3. Leave Require Approval Workflow enabled (recommended for initial setup)
  4. Click Save Changes
  5. Assign the create_agent and/or create_sme tool to the agents that should be able to create new agents
  6. Monitor incoming requests on the Approvals page

Best Practices​

  • Start with approval enabled - Enable the approval workflow while getting familiar with the feature, then consider disabling it once you trust the patterns
  • Use templates - Create templates for common agent types to ensure consistency and reduce risk
  • Set conservative rate limits - Start with low limits and increase as needed
  • Monitor the logs - Regularly check the creation logs for unexpected patterns
  • Keep the safety threshold at 30% - The default threshold balances safety with usability
  • Assign tools carefully - Only give the create_agent tool to orchestrator-level agents that genuinely need it