Sub-Agents

Build specialized agents that work together to handle complex tasks.

What are Sub-Agents?

Sub-agents are specialized AI agents that your main agent can delegate tasks to. This architecture allows you to:

• Specialize: Create experts in specific domains
• Scale: Handle more complex workflows
• Maintain: Update individual agents without affecting others
• Optimize: Use different models for different tasks

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    Multi-Agent System                        │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│                    ┌──────────────────┐                     │
│                    │   Main Agent     │                     │
│                    │  (Orchestrator)  │                     │
│                    └────────┬─────────┘                     │
│                             │                                │
│           ┌─────────────────┼─────────────────┐             │
│           │                 │                 │             │
│           ▼                 ▼                 ▼             │
│   ┌───────────────┐ ┌───────────────┐ ┌───────────────┐    │
│   │  Technical    │ │   Billing     │ │    Sales      │    │
│   │  Support      │ │   Support     │ │   Assistant   │    │
│   │  Agent        │ │   Agent       │ │   Agent       │    │
│   └───────────────┘ └───────────────┘ └───────────────┘    │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Use Cases

1. Customer Support Routing

Customer: "I'm having trouble with my API integration"
         │
         ▼
Main Agent: Identifies as technical issue
         │
         ▼
Technical Support Agent: Handles API troubleshooting

2. Multi-Step Workflows

Customer: "I want to upgrade my plan and add a user"
         │
         ▼
Main Agent: Identifies two tasks
         │
         ├──► Billing Agent: Process upgrade
         │
         └──► User Management Agent: Add team member

3. Language Specialization

Customer: "Bonjour, j'ai besoin d'aide"
         │
         ▼
Main Agent: Detects French
         │
         ▼
French Support Agent: Handles in French

Creating Sub-Agents

Step 1: Create Specialized Agent

Create a new agent for each specialization:

1. Go to Agents > Create Agent
2. Configure for specific domain:

markdownCopy
# Technical Support Agent

You are a technical support specialist for CloudAPI.

You handle:
- API integration issues
- SDK troubleshooting
- Authentication problems
- Rate limiting questions

You have access to:
- Technical documentation
- API reference
- Known issues database
- Error code lookup tool

Always provide code examples when relevant.

Step 2: Configure as Sub-Agent

Mark the agent as a sub-agent:

1. Go to agent settings
2. Enable "Available as Sub-Agent"
3. Set visibility (which agents can call it)

Step 3: Connect to Main Agent

Add sub-agents to your orchestrator:

1. Open your main agent
2. Go to Sub-Agents tab
3. Click Add Sub-Agent
4. Select your specialized agents
5. Configure routing rules

Routing Strategies

Intent-Based Routing

Route based on detected intent:

jsonCopy
{
  "routing_rules": [
    {
      "intent": "technical_issue",
      "sub_agent": "technical-support",
      "keywords": ["api", "error", "bug", "integration", "code"]
    },
    {
      "intent": "billing_inquiry",
      "sub_agent": "billing-support",
      "keywords": ["invoice", "payment", "charge", "refund", "plan"]
    },
    {
      "intent": "sales_question",
      "sub_agent": "sales-assistant",
      "keywords": ["pricing", "demo", "enterprise", "features"]
    }
  ]
}

Explicit Routing

Let the main agent decide:

markdownCopy
# Main Agent Instructions

When you identify a specialized need, delegate to the appropriate agent:

- Technical issues → @technical-support
- Billing questions → @billing-support
- Sales inquiries → @sales-assistant

Always inform the user when transferring:
"Let me connect you with our technical specialist..."

Fallback Routing

Handle unmatched queries:

jsonCopy
{
  "fallback": {
    "action": "handle_self",
    "message": "I'll help you directly with this."
  }
}

Context Passing

Conversation Context

Pass relevant history to sub-agents:

jsonCopy
{
  "context_passing": {
    "include_history": true,
    "max_messages": 5,
    "include_metadata": true
  }
}

Custom Context

Add specific information:

jsonCopy
{
  "context": {
    "customer_tier": "enterprise",
    "account_age_days": 365,
    "previous_issues": ["billing-2024-001", "tech-2024-015"]
  }
}

Response Handling

Direct Response

Sub-agent responds directly to user:

User → Main Agent → Sub-Agent → User
                    (responds directly)

Filtered Response

Main agent reviews before sending:

User → Main Agent → Sub-Agent → Main Agent → User
                    (filters/enhances response)

Configuration

jsonCopy
{
  "response_handling": {
    "mode": "filtered",
    "filter_rules": {
      "remove_internal_links": true,
      "add_branding": true,
      "check_tone": true
    }
  }
}

Orchestration Patterns

Sequential Processing

Handle tasks one at a time:

Task 1 → Agent A → Complete
                      │
                      ▼
Task 2 → Agent B → Complete
                      │
                      ▼
Return combined result

Parallel Processing

Handle tasks simultaneously:

         ┌→ Agent A → Result A ─┐
         │                      │
Tasks ───┼→ Agent B → Result B ─┼→ Combined Result
         │                      │
         └→ Agent C → Result C ─┘

Configuration

jsonCopy
{
  "orchestration": {
    "mode": "parallel",
    "timeout_ms": 30000,
    "fail_strategy": "continue_others"
  }
}

Best Practices

1. Clear Boundaries

Each sub-agent should have a well-defined scope:

2. Consistent Handoffs

Make transitions smooth:

markdownCopy
Main Agent: "I'll connect you with our billing team who can help with that invoice question."

Billing Agent: "Hi! I'm here to help with your billing inquiry. I can see you have a question about your recent invoice..."

3. Fallback Handling

Always have a fallback:

jsonCopy
{
  "fallback_behavior": {
    "no_match": "handle_self",
    "sub_agent_error": "escalate_to_human",
    "timeout": "retry_once"
  }
}

4. Monitor Performance

Track sub-agent effectiveness:

Next Steps

• Deployment - Deploy your multi-agent system
• Workflows - Connect agents to workflows