Skip to main content

Orchestration

The Orchestration panel lets you define rules for agent behavior, escalation paths, and coordination between agents.

Overview​

Orchestration rules control:

  • When to escalate to other agents
  • Handoff conditions
  • Approval requirements
  • Workflow triggers

Escalation Rules​

Define Escalation Paths​

const rules = await client.orchestration.createRules('agent-123', {
  escalation: {
    // Escalate after 3 failed attempts
    maxAttempts: 3,

    // Escalate for specific intents
    intents: ['billing_dispute', 'legal_question', 'security_incident'],

    // Target agent for escalation
    targetAgentId: 'agent-supervisor',

    // Or escalate to human
    escalateToHuman: true
  }
});

Handoff Rules​

Automatic Handoffs​

const rules = await client.orchestration.createRules('agent-123', {
  handoffs: [
    {
      condition: 'intent == "technical_support"',
      targetAgentId: 'agent-tech-support'
    },
    {
      condition: 'intent == "billing"',
      targetAgentId: 'agent-billing'
    },
    {
      condition: 'sentiment < -0.5',
      targetAgentId: 'agent-retention'
    }
  ]
});

Approval Requirements​

Require Approval for Sensitive Actions​

const rules = await client.orchestration.createRules('agent-123', {
  approvals: {
    required: true,
    actions: [
      'issue_refund',
      'modify_account',
      'access_sensitive_data'
    ],
    approvers: ['user-admin-123', 'user-manager-456']
  }
});

Coordination Rules​

Multi-Agent Workflows​

const workflow = await client.workflows.create({
  name: 'Support Escalation',
  steps: [
    {
      agent: 'agent-triage',
      condition: 'start',
      outputs: ['category', 'priority']
    },
    {
      agent: 'agent-specialist',
      condition: 'category == "technical"',
      inputs: ['category', 'context']
    },
    {
      agent: 'agent-supervisor',
      condition: 'priority == "critical"',
      inputs: ['all']
    }
  ]
});

Best Practices​

  1. Clear conditions: Make escalation triggers explicit
  2. Fallback paths: Always have a fallback (usually human)
  3. Test thoroughly: Verify rules work as expected
  4. Monitor: Track escalation rates to tune rules

Next: Learn about Publishing your templates.