router

Router Pattern

A classifier agent examines the user's input, determines intent, and routes to the appropriate specialist agent. Each specialist handles a specific domain with tailored system prompts.

When to Use

• Customer support: Route tickets to billing, technical, or general support
• Help desks: Dispatch queries to the right team based on topic
• Any scenario where the response strategy depends on input classification

Architecture

graph LR
    Input[User Input] --> Router[Router Agent<br/>classifier]
    Router -->|BILLING| Billing[Billing Specialist]
    Router -->|TECHNICAL| Technical[Technical Specialist]
    Router -->|GENERAL| General[General Specialist]
    Billing --> Output[Response]
    Technical --> Output
    General --> Output

Loading

How It Works

1. Router agent receives the user's message and classifies intent into one of three categories: BILLING, TECHNICAL, or GENERAL
2. A handoff event is emitted indicating which specialist was selected and why
3. The matched specialist agent processes the original input with domain-specific expertise
4. The specialist's streamed response becomes the final output

Event Flow

agent_start  {agent: "router", role: "classifier"}
chunk        {agent: "router", content: "BILLING"}
agent_end    {agent: "router", ...}
handoff      {from: "router", to: "billing", reason: "billing intent detected"}
agent_start  {agent: "billing", role: "specialist"}
chunk        {agent: "billing", content: "I can help with your invoice..."}
agent_end    {agent: "billing", ...}
done         {totalUsage: {...}}

Specialists

Agent	Handles
billing	Invoices, payments, refunds, subscriptions, pricing
technical	Bugs, crashes, errors, performance, feature help
general	Business hours, company info, accounts, everything else

If the router's output does not match a known category, it defaults to GENERAL.

Tradeoffs

• Fast: Only two LLM calls (classify + respond), minimal latency overhead
• Simple: Easy to add new specialists by extending the category list
• Rigid: Classification is single-label; ambiguous inputs may misroute
• No fallback loop: If the specialist gives a poor answer, there is no retry mechanism