A single agent can't do everything well. Multi-agent systems use a supervisor that routes tasks to specialist agents: a research agent for information gathering, a planner for scheduling, an executor for actions. Each specialist is narrow and effective.
Supervisor Agent Pattern
Code Snippet
Supervisor: "User wants to book a flight. I should ask the research agent to find flights."
↓
Research Agent: "Finds available flights based on user requirements"
↓
Supervisor: "I have flight options. I'll ask the booking agent to complete the transaction."
↓
Booking Agent: "Completes the booking"
Implementing Multi-Agent with LangGraph
Code Snippet
from langgraph.graph import StateGraph, START, END
from langchain_openai import ChatOpenAI
# Define agents
research_agent = create_research_agent()
booking_agent = create_booking_agent()
# Define supervisor
def supervisor(state):
# Decide which agent to call
if "flight" in state["request"]:
return "research"
elif "book" in state["request"]:
return "booking"
return "end"
# Build graph
workflow = StateGraph(State)
workflow.add_node("supervisor", supervisor)
workflow.add_node("research", research_agent)
workflow.add_node("booking", booking_agent)
# Edges
workflow.add_conditional_edges("supervisor", supervisor)
workflow.add_edge("research", "booking")
workflow.add_edge("booking", END)
workflow.set_entry_point("supervisor")
graph = workflow.compile()
Shared Memory
Code Snippet
state = {
"user_request": "Book a flight to Paris",
"flights": [], # Populated by research agent
"booking_id": None # Populated by booking agent
}
result = graph.invoke(state)
Each agent reads and writes to shared state, enabling multi-turn, context-aware workflows.
Conclusion
Multi-agent systems scale to complex tasks by using specialized agents. The supervisor orchestrates delegation. Shared memory enables context flow between agents. Building multi-agent systems requires understanding state management and agent coordination. Next: deploying agents to production with proper error handling and monitoring.