A router selects which specialist handles an incoming request. In a multi-agent system, it chooses the next agent; in agentic RAG, it can direct a query to a specialized retrieval tool or knowledge store, such as vector search, graph traversal, or an external API. section 7.7.4 section 3.5
Routing fits requests that fall into mutually exclusive categories. Unlike a query-planning agent, which decomposes a question into retrieval subtasks, a router selects the appropriate destination. Its main risks are misclassification and unhandled routes. section 3.5 section 7.7
Customer support routing
A support workflow classifies each question as sales, technical support, or general. A routing function reads the current user message and any retained conversation messages from workflow state, asks a model to classify the question, and returns the branch name. The selected specialist then handles the request. section 7.9.3
Listing 7.4 Building the LangGraph workflow for the router pattern
def build_graph(self):
workflow = StateGraph(AgentState) #1
workflow.add_node("router", self.router_node) #2
workflow.add_node("sales_agent", self.create_agent_node("You are a
helpful sales assistant."))
workflow.add_node("tech_support_agent", self.create_agent_node("You
are a helpful technical support assistant."))
workflow.add_node("general_agent", self.create_agent_node("You are a
helpful general assistant."))
workflow.set_entry_point("router") #3
workflow.add_conditional_edges( #4
"router",
self.route_question, #5
{
"sales": "sales_agent",
"tech_support": "tech_support_agent",
"general": "general_agent",
},
)
workflow.add_edge("sales_agent", END)
workflow.add_edge("tech_support_agent", END)
workflow.add_edge("general_agent", END)
return workflow.compile()
The LangGraph listing shows the router as the graph entry point and maps classification results to specialist agent nodes through conditional edges. section 7.9.3
A production routing decision should include a route name, confidence score, and short reason. Apply a confidence threshold and send ambiguous or unknown requests to an explicit fallback or user clarification path rather than silently defaulting to the general agent. Choose the most constrained coordination pattern that meets the task’s needs: greater flexibility can reduce predictability and debuggability. section 7.7.4 section 7.9.3 section 7.7

