# Multi-Agent: Handoff Pattern Ref: - https://docs.langchain.com/oss/python/langchain/multi-agent - https://docs.langchain.com/oss/python/langgraph/use-subgraphs The handoff pattern allows agents to directly pass control to each other. **Characteristics:** - Decentralized control flow - Active agent changes during conversation - Good for multi-domain conversations with specialists **Example:** Router hands off to either a math agent or a translation agent based on the request. ## Setup Configure `.env` before running. See `.env.sample`. ```python import rich from dotenv import load_dotenv load_dotenv() ``` True ## Define State and Agents Create specialized agents that can hand off to each other using `Command`. ```python from typing import Annotated from langchain_anthropic import ChatAnthropic from langchain_core.messages import AIMessage, BaseMessage, HumanMessage from langgraph.graph import END, START, StateGraph from langgraph.graph.message import add_messages from langgraph.types import Command from pydantic import BaseModel class State(BaseModel): """Shared state between agents.""" messages: Annotated[list[BaseMessage], add_messages] current_agent: str = "router" model = ChatAnthropic(model="claude-sonnet-4-5-20250929") ``` ## Create Router Agent The router determines which specialist should handle the request. ```python def router_agent(state: State) -> Command[str]: """Route to the appropriate specialist based on the query.""" messages = state.messages if not messages: return Command(goto=END) last_message = messages[-1] # If last message is from an agent (AIMessage), we're done if isinstance(last_message, AIMessage): return Command(goto=END) # Ask the model to decide which agent should handle this response = model.invoke( [ { "role": "system", "content": """You are a router. Based on the user's message, decide which specialist should handle it. Reply with ONLY one word: 'math', 'translate', or 'done'. - math: Calculations, arithmetic, math problems - translate: Translation between languages - done: Conversation is complete or greeting/goodbye""", }, {"role": "user", "content": str(last_message.content)}, ] ) decision = str(response.content).strip().lower() rich.print(f"[dim]Router decision: {decision}[/dim]") if "math" in decision: return Command(goto="math_agent", update={"current_agent": "math"}) elif "translate" in decision: return Command(goto="translate_agent", update={"current_agent": "translate"}) else: return Command(goto=END) ``` ## Create Specialist Agents Each specialist handles their domain and can hand back to the router. ```python def _format_messages(messages: list[BaseMessage]) -> list[dict[str, str]]: """Format messages for model invocation.""" return [ {"role": "user" if isinstance(m, HumanMessage) else "assistant", "content": str(m.content)} for m in messages ] def math_agent(state: State) -> Command[str]: """Handle math calculations.""" response = model.invoke( [ { "role": "system", "content": """You are a math specialist. Solve calculations and math problems. Show your work step by step, then give the final answer.""", }, *_format_messages(state.messages), ] ) return Command( goto="router", update={"messages": [AIMessage(content=str(response.content))], "current_agent": "router"}, ) def translate_agent(state: State) -> Command[str]: """Handle translations.""" response = model.invoke( [ { "role": "system", "content": """You are a translation specialist. Translate text between languages. If the target language is not specified, translate to English. Provide the translation directly.""", }, *_format_messages(state.messages), ] ) return Command( goto="router", update={"messages": [AIMessage(content=str(response.content))], "current_agent": "router"}, ) ``` ## Build the Graph Connect the agents in a graph structure. ```python builder = StateGraph(State) # Add nodes builder.add_node("router", router_agent) builder.add_node("math_agent", math_agent) builder.add_node("translate_agent", translate_agent) # Start with router builder.add_edge(START, "router") # Compile graph = builder.compile() rich.print("Graph compiled with nodes: router, math_agent, translate_agent") ``` Graph compiled with nodes: router, math_agent, translate_agent ## Test: Math Question ```python result = graph.invoke({"messages": [HumanMessage(content="What is 25 * 4 + 15?")]}) # type: ignore[arg-type] rich.print("Final response:") rich.print(result["messages"][-1].content) ``` Router decision: math Final response: I'll solve this step by step, following the order of operations (PEMDAS/BODMAS). **Step 1:** First, perform the multiplication 25 × 4 = 100 **Step 2:** Then, perform the addition 100 + 15 = 115 **Final Answer:** 115 ## Test: Translation Question ```python query = "Translate 'Hello, how are you?' to Japanese" result = graph.invoke({"messages": [HumanMessage(content=query)]}) # type: ignore[arg-type] rich.print("Final response:") rich.print(result["messages"][-1].content) ``` Router decision: translate Final response: こんにちは、お元気ですか？ ## Stream Execution Stream to see the handoff flow between agents. ```python for chunk in graph.stream({"messages": [HumanMessage(content="What is 100 divided by 8?")]}): # type: ignore[arg-type] rich.print("chunk =", chunk) ``` Router decision: math chunk = {'router': {'current_agent': 'math'}} chunk = { 'math_agent': { 'messages': [ AIMessage( content='I need to calculate 100 divided by 8.\n\n**Step-by-step calculation:**\n\n100 ÷ 8\n\nI can work this out by finding how many times 8 goes into 100:\n- 8 × 10 = 80\n- 8 × 12 = 96\n- 8 × 12.5 = 100\n\nSo 8 goes into 100 exactly 12.5 times.\n\nAlternatively, I can express this as:\n100 ÷ 8 = 100/8 = 25/2 = 12.5\n\nOr as a mixed number:\n100 ÷ 8 = 12 remainder 4 = 12 4/8 = 12 1/2\n\n**Final answer: 12.5** (or 12½ or 25/2)', additional_kwargs={}, response_metadata={}, id='a258a5eb-8e7d-4ca1-9eda-e50382249bbd' ) ], 'current_agent': 'router' } } chunk = {'router': None}