LangGraphにおける階層型マルチエージェントシステム：Supervisor、サブグラフ、ステートマシン

from typing import List, Annotated, TypedDict, Union
from langchain_core.messages import BaseMessage, HumanMessage, AIMessage

class AgentState(TypedDict):
    """
    Represents the state of our multi-agent system.
    This state is shared across all agents and subgraphs.
    """
    messages: Annotated[List[BaseMessage], operator.add]
    next_action: str # The next action the supervisor decides to take
    task: str # The initial task given to the supervisor
    research_results: Annotated[List[str], operator.add]
    code_output: str
    verification_status: str
    error_message: str # For error recovery
    iterations: int # To prevent infinite loops

# Example of how to initialize the state
initial_state = AgentState(
    messages=[HumanMessage(content="Initial task description.")],
    next_action="supervisor_decision",
    task="Initial task description.",
    research_results=[],
    code_output="",
    verification_status="pending",
    error_message="",
    iterations=0
)

import operator
from langgraph.graph import StateGraph, END, START
from langgraph.checkpoint.sqlite import SqliteSaver
from langgraph.prebuilt import ToolNode
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.messages import BaseMessage, HumanMessage, AIMessage
from typing import List, Annotated, TypedDict, Union

# Assume AgentState is defined as above

class SupervisorAgent:
    def __init__(self, llm):
        self.llm = llm
        self.prompt = ChatPromptTemplate.from_messages([
            ("system", "You are a highly intelligent supervisor agent. Your goal is to orchestrate a team of specialized agents to complete a given task. Based on the current state and messages, decide the next action. Available actions: {actions}. If the task is complete, respond with 'FINISH'. If an error occurred, respond with 'ERROR_RECOVERY'."),
            ("user", "{messages}")
        ])
        self.router = self.prompt | self.llm.bind_tools(
            tools=[
                {"name": "research", "description": "Delegate to the research agent."},
                {"name": "code_execution", "description": "Delegate to the code execution agent."},
                {"name": "verification", "description": "Delegate to the verification agent."},
                {"name": "finish", "description": "The task is complete."},
                {"name": "error_recovery", "description": "An error occurred, attempt recovery."}
            ]
        )

    def route_agent(self, state: AgentState) -> str:
        """
        Routes the execution based on the supervisor's decision.
        """
        print(f"---SUPERVISOR DECIDING--- Iteration: {state['iterations']}")
        state['iterations'] += 1
        if state['iterations'] > 10: # Safety break
            return "FINISH" # Or "ERROR_RECOVERY"

        response = self.router.invoke({"messages": state["messages"], "actions": ["research", "code_execution", "verification", "FINISH", "ERROR_RECOVERY"]})
        tool_calls = response.tool_calls
        if tool_calls:
            action = tool_calls[0]['name']
            print(f"Supervisor chose: {action}")
            return action
        else:
            # If LLM doesn't call a tool, it might be trying to finish or error
            content = response.content.strip().upper()
            if "FINISH" in content:
                print("Supervisor chose: FINISH")
                return "FINISH"
            elif "ERROR_RECOVERY" in content:
                print("Supervisor chose: ERROR_RECOVERY")
                return "ERROR_RECOVERY"
            else:
                print(f"Supervisor made an ambiguous decision: {content}. Defaulting to FINISH.")
                return "FINISH" # Fallback

from langchain_community.tools import DuckDuckGoSearchRun

class ResearchAgent:
    def __init__(self, llm):
        self.llm = llm
        self.search_tool = DuckDuckGoSearchRun()
        self.prompt = ChatPromptTemplate.from_messages([
            ("system", "You are a research assistant. Use the provided search tool to gather information relevant to the user's task. Summarize your findings concisely. If you have enough information, respond with 'DONE'."),
            ("user", "{messages}")
        ])
        self.research_chain = self.prompt | self.llm.bind_tools(tools=[self.search_tool])

    def research_node(self, state: AgentState) -> AgentState:
        print("---RESEARCH AGENT---")
        # Extract the latest human message as the query
        query = state["messages"][-1].content if state["messages"] else state["task"]
        response = self.research_chain.invoke({"messages": state["messages"]})

        tool_calls = response.tool_calls
        if tool_calls:
            # Assuming the research agent will call the search tool
            tool_output = self.search_tool.invoke(tool_calls[0]['args']['query'])
            state["research_results"].append(f"Search result for '{tool_calls[0]['args']['query']}': {tool_output}")
            state["messages"].append(AIMessage(content=f"Performed search. Results added to state. Current research: {tool_output[:100]}..."))
            state["next_action"] = "supervisor_decision" # Return control to supervisor
        else:
            # If no tool call, it means the research agent might be done or summarizing
            state["research_results"].append(response.content)
            state["messages"].append(AIMessage(content=f"Research summary: {response.content}"))
            state["next_action"] = "supervisor_decision" # Return control to supervisor

        return state

# Build the research subgraph
def create_research_subgraph(llm):
    research_agent = ResearchAgent(llm)
    research_graph = StateGraph(AgentState)
    research_graph.add_node("research_node", research_agent.research_node)
    research_graph.add_edge(START, "research_node")
    research_graph.add_edge("research_node", END) # Research node always returns to supervisor
    return research_graph.compile()

from langchain_community.tools import PythonREPLTool

class CodeExecutionAgent:
    def __init__(self, llm):
        self.llm = llm
        self.python_repl = PythonREPLTool()
        self.prompt = ChatPromptTemplate.from_messages([
            ("system", "You are a coding assistant. Execute Python code to solve the task. If an error occurs, try to fix it. Respond with 'DONE' when the code is successfully executed and verified."),
            ("user", "{messages}")
        ])
        self.code_chain = self.prompt | self.llm.bind_tools(tools=[self.python_repl])

    def execute_code_node(self, state: AgentState) -> AgentState:
        print("---CODE EXECUTION AGENT---")
        try:
            response = self.code_chain.invoke({"messages": state["messages"]})
            tool_calls = response.tool_calls
            if tool_calls:
                # Assuming the code agent will call the python_repl tool
                code_to_execute = tool_calls[0]['args']['code']
                print(f"Executing code:\n{code_to_execute}")
                tool_output = self.python_repl.invoke({"code": code_to_execute})
                state["code_output"] = tool_output
                state["messages"].append(AIMessage(content=f"Code executed. Output: {tool_output}"))
                state["next_action"] = "supervisor_decision"
            else:
                state["messages"].append(AIMessage(content=f"Code agent response: {response.content}"))
                state["next_action"] = "supervisor_decision" # If no tool call, it might be done or summarizing
        except Exception as e:
            state["error_message"] = f"Code execution failed: {str(e)}"
            state["messages"].append(AIMessage(content=f"Code execution failed: {str(e)}. Attempting error recovery."))
            state["next_action"] = "error_recovery" # Signal supervisor for recovery
        return state

# Build the code execution subgraph
def create_code_execution_subgraph(llm):
    code_agent = CodeExecutionAgent(llm)
    code_graph = StateGraph(AgentState)
    code_graph.add_node("execute_code_node", code_agent.execute_code_node)
    code_graph.add_edge(START, "execute_code_node")
    code_graph.add_edge("execute_code_node", END)
    return code_graph.compile()

class VerificationAgent:
    def __init__(self, llm):
        self.llm = llm
        self.prompt = ChatPromptTemplate.from_messages([
            ("system", "You are a verification agent. Your task is to verify the results of previous steps, especially code execution. Identify any discrepancies or errors. Respond with 'VERIFIED' if successful, or 'NEEDS_REVISION' if issues are found."),
            ("user", "{messages}")
        ])
        self.verify_chain = self.prompt | self.llm

    def verify_node(self, state: AgentState) -> AgentState:
        print("---VERIFICATION AGENT---")
        response = self.verify_chain.invoke({"messages": state["messages"]})
        verification_result = response.content.strip().upper()

        if "VERIFIED" in verification_result:
            state["verification_status"] = "verified"
            state["messages"].append(AIMessage(content="Verification successful."))
        else:
            state["verification_status"] = "needs_revision"
            state["messages"].append(AIMessage(content=f"Verification failed: {response.content}. Needs revision."))
            state["error_message"] = f"Verification failed: {response.content}" # Set error for potential recovery

        state["next_action"] = "supervisor_decision"
        return state

# Build the verification subgraph
def create_verification_subgraph(llm):
    verification_agent = VerificationAgent(llm)
    verification_graph = StateGraph(AgentState)
    verification_graph.add_node("verify_node", verification_agent.verify_node)
    verification_graph.add_edge(START, "verify_node")
    verification_graph.add_edge("verify_node", END)
    return verification_graph.compile()

from langgraph.graph import StateGraph, END, START
from langgraph.checkpoint.sqlite import SqliteSaver
import os

# Initialize LLM (e.g., OpenAI)
# Ensure OPENAI_API_KEY is set in environment variables
llm = ChatOpenAI(model="gpt-4o", temperature=0)

# Create agents and subgraphs
supervisor_agent = SupervisorAgent(llm)
research_subgraph = create_research_subgraph(llm)
code_execution_subgraph = create_code_execution_subgraph(llm)
verification_subgraph = create_verification_subgraph(llm)

# Define the main graph
workflow = StateGraph(AgentState)

# Add nodes for supervisor and subgraphs
workflow.add_node("supervisor_decision", supervisor_agent.route_agent)
workflow.add_node("research", research_subgraph)
workflow.add_node("code_execution", code_execution_subgraph)
workflow.add_node("verification", verification_subgraph)

# Define conditional edges for the supervisor
workflow.add_conditional_edges(
    "supervisor_decision",
    lambda state: state["next_action"], # The supervisor's output determines the next node
    {
        "research": "research",
        "code_execution": "code_execution",
        "verification": "verification",
        "FINISH": END,
        "ERROR_RECOVERY": "error_recovery_node" # Placeholder for error recovery
    }
)

# Add a dedicated error recovery node (can be another agent or a human-in-the-loop)
def error_recovery_node(state: AgentState) -> AgentState:
    print(f"---ERROR RECOVERY--- Error: {state['error_message']}")
    # Here, you could implement more sophisticated recovery logic:
    # - Summarize error and ask supervisor to re-plan
    # - Notify human operator
    # - Attempt a retry with modified parameters
    state["messages"].append(AIMessage(content=f"Attempting error recovery for: {state['error_message']}"))
    state["error_message"] = "" # Clear error after handling attempt
    state["next_action"] = "supervisor_decision" # Return to supervisor for re-evaluation
    return state

workflow.add_node("error_recovery_node", error_recovery_node)
workflow.add_edge("error_recovery_node", "supervisor_decision")

# Edges from subgraphs back to supervisor
workflow.add_edge("research", "supervisor_decision")
workflow.add_edge("code_execution", "supervisor_decision")
workflow.add_edge("verification", "supervisor_decision")

# Set the entry point
workflow.set_entry_point("supervisor_decision")

# Compile the graph with memory
memory = SqliteSaver.from_conn_string(":memory:") # Use a file path for persistence
app = workflow.compile(checkpointer=memory)

# Example execution
config = {"configurable": {"thread_id": "user-task-123"}}
initial_task = "Research the capital of France, then write and execute Python code to calculate 2+2, and verify the result."
initial_messages = [HumanMessage(content=initial_task)]

# First run
print("\n--- Initial Run ---")
for s in app.stream({"messages": initial_messages, "task": initial_task, "iterations": 0}, config=config):
    if "__end__" not in s:
        print(s)
        print("---")

# Retrieve final state
final_state = app.get_state(config)
print("\n--- Final State ---")
print(final_state.values)

# Demonstrate human-in-the-loop (HITL) and rollback
# Imagine a human reviews the state and finds an issue, then modifies it.
# This is where the checkpointer is crucial.

# Let's simulate a human intervention after some steps
# We can load a specific checkpoint or modify the current state
# For demonstration, we'll just modify the current state and re-run
# In a real scenario, a human might edit the state via a UI and then resume.

# Simulate an error in code execution and a human fixing it
# We'll manually set the state to simulate an error and then a fix
# For a real HITL, you'd pause, present the state, allow edits, then resume.

# Let's assume the code execution failed and we want to retry
# We can manually set the state to trigger error recovery or a specific action
# This is a simplified example; a real HITL would involve UI interaction.

# Example of loading a specific checkpoint (if we had multiple saved)
# from langgraph.checkpoint.base import Checkpoint
# checkpoint: Checkpoint = memory.get(config)
# print(f"Loaded checkpoint: {checkpoint}")

# For this example, we'll just re-run from the current state,
# but if we wanted to rollback, we'd load an earlier checkpoint.

# Let's simulate a human reviewing the research and adding more context
print("\n--- Simulating Human Intervention (Adding more research context) ---")
current_state = app.get_state(config).values
current_state["research_results"].append("Human added: Paris is also known as the 'City of Light'.")
current_state["messages"].append(HumanMessage(content="Human review: Added more context about Paris. Please proceed."))
current_state["next_action"] = "supervisor_decision" # Force supervisor to re-evaluate

# Resume from the modified state
print("\n--- Resuming after Human Intervention ---")
for s in app.stream(current_state, config=config):
    if "__end__" not in s:
        print(s)
        print("---")

final_state_after_hitl = app.get_state(config)
print("\n--- Final State After HITL ---")
print(final_state_after_hitl.values)