How to use the AgentExecutor.from_agent_and_tools with create_react_agent as a node in Langgraph ?

[AkashBais]

I am trying to use AgentExecutor.from_agent_and_tools with create_react_agent as the agent inside a LangGraph node, but the graph state gets automatically modified before it's passed to the AgentExecutor object.

I create a agent node with the code below

def initializes_agent(llm,
                 tools: List[tool],
                 system_message: str,
                 team_members: List[str],
                 working_directory: str) -> AgentExecutor:
    """
    Creates an AI agent with specified tools and configuration for a data analysis team.

    Args:
        llm: The language model to be used by the agent.
        tools (List[tool]): A list of tools available to the agent.
        system_message (str): The system message defining the agent's role and instructions.
        team_members (List[str]): A list of team members' names.
        working_directory (str): The working directory for the agent.

    Returns:
        AgentExecutor: The created agent executor with the specified configuration.

    Raises:
        Exception: If any error occurs during the agent creation process.
    """
    try:
        logger.info("Initializing agent creation")
        
        # Ensure the list_directory_contents tool is included in the tools list
        if list_directory_contents not in tools:
            tools.append(list_directory_contents)

        # Prepare tool names and team members as comma-separated strings
        tool_names = ', '.join([tool.name for tool in tools])
        team_members_str = ', '.join([str(member) for member in team_members])

        # Get the initial contents of the working directory
        # initial_directory_contents = list_directory_contents(path=working_directory)
        initial_directory_contents = list_directory_contents(working_directory)

        # Define the system prompt for the agent
        system_prompt = (
            "You are a specialized AI assistant in a data analysis team. "
            "Your role is to complete specific tasks in the research process. "
            "Use the provided tools to make progress on your task. "
            "If you can't fully complete a task, explain what you've done and what's needed next. "
            "Always aim for accurate and clear outputs. "
            f"You have access to the following tools: {tool_names}. "
            f"Your specific role: {system_message}\n"
            "Work autonomously according to your specialty, using the tools available to you. "
            "Do not ask for clarification. "
            "Your other team members (and other teams) will collaborate with you based on their specialties. "
            f"You are chosen for a reason! You are one of the following team members: {team_members_str}.\n"
            f"The initial contents of your working directory are:\n{initial_directory_contents}\n"
            "Use the ListDirectoryContents tool to check for updates in the directory contents when needed."
        )

        # Create the prompt template for the agent
        prompt = ChatPromptTemplate.from_messages(
            [
                ('system', system_prompt),
                MessagesPlaceholder(variable_name="messages"),
                ("ai", "hypothesis:{hypothesis}"),
                ("ai", "process:{process}"),
                ("ai", "process_decision:{process_decision}"),
                ("ai", "visualization_state:{visualization_state}"),
                ("ai", "search_state:{search_state}"),
                ("ai", "code_state:{code_state}"),
                ("ai", "report_section:{report_section}"),
                ("ai", "quality_review:{quality_review}"),
                ("ai", "needs_revision:{needs_revision}"),
                MessagesPlaceholder(variable_name="agent_scratchpad"),
            ]
        )

        # Create the agent with the specified model, tools, and prompt
        agent = create_react_agent(
            model=llm,
            tools=tools,
            state_modifier=prompt,
            debug=True,
        )

        logger.info("Agent creation completed successfully")
        return AgentExecutor.from_agent_and_tools(agent=agent,
                                                  tools=tools,
                                                  verbose=True)
    
    except Exception as e:
        logger.error(f"Error in agent creation: {str(e)}")
        raise e

This agent is then called within the below method

def agent_node(agent: AgentExecutor, 
               state: State, 
               name: str) -> State:
    """
    Processes a given agent node, updates the state based on the agent's response, and handles any exceptions.

    Args:
        agent (AgentExecutor): The agent to be invoked.
        state (State): The current state of the system.
        name (str): The name of the agent being processed.

    Returns:
        State: The updated state after processing the agent.

    Raises:
        Exception: If any error occurs during the execution of the agent node.
    """
    try:
        logger.info(f"Processing {name} agent")
        
        logger.info(f"State: {state}")
        
        # Log state before invocation
        logger.debug(f"State before invoking {name}: {state}")  
        
        # Invoke the agent with the current state
        response = agent.invoke(state)
        
        # Extract the content from the response
        content = response['output'] if isinstance(response, dict) and 'output' in response else str(response)
        
        # Create an AIMessage with the content and agent name
        ai_message = AIMessage(content=content, name=name)
        
        # Append the AIMessage to the state's messages
        state['messages'].append(ai_message)
        state['sender'] = name
        
        logger.info("Updating state")
        
        # Update the state based on the agent's name
        if name == "hypothesis_agent" and not state["hypothesis"]:
            state["hypothesis"] = ai_message
            
        elif name == "process_agent":
            state["process_decision"] = ai_message
            
        elif name == "visualization_agent":
            state["visualization_state"] = ai_message
            
        elif name == "search_agent":
            state["search_state"] = ai_message
            
        elif name == "code_agent":
            state["code_state"] = ai_message
            
        elif name == "report_agent":
            state["report_section"] = ai_message
            
        elif name == "quality_review":
            state["quality_review"] = ai_message
            state["needs_revision"] = "revision needed" in response.lower()
        
        logger.info(f"Processing {name} agent completed")
        
        return state
    
    except Exception as e:
        logger.error(f"Error while executing {name} node: {str(e)}", exc_info=True)
        error_message = AIMessage(content=f"Error while executing {name} node: {str(e)}", name=name)
        
        # Return an error state
        return create_error_state(state=state, error_message=error_message, name=name, error_type=type(e))

I am invoking the graph as follows

graph.invoke( input =  {"messages" : [message],
                      "visualization_state" : 'Initial visualization_state content',# AIMessage(content= 'Initial visualization_state content'),
                      "report_section" : 'Initial report_section content',#AIMessage(content= 'Initial report_section content'),
                      "needs_revision" : False,
                      "agent_scratchpad" : 'Initial agent_scratchpad content',#AIMessage(content= 'Initial agent_scratchpad content'),
                      "process" : 'Initial process content',#AIMessage(content= 'Initial process content'), 
                      "search_state" : 'Initial search_state content',#AIMessage(content= 'Initial search_state content'),
                      "code_state" : 'Initial code_state content',#AIMessage(content= 'Initial code_state content'),
                      "hypothesis" : 'Initial hypothesis content',#AIMessage(content= 'Initial hypothesis content'),
                      "quality_review" : 'Initial quality_review content',#AIMessage(content= 'Initial quality_review content'),
                      "process_decision" : 'Initial process_decision content',#AIMessage(content= 'Initial process_decision content'),
                       "agent_scratchpad": [],
              },
            config = {"configurable": {"thread_id": "42"}, "recursion_limit": 30} ,
            
            )

I get the following error

2024-12-24 16:23:04,244 - core.Logger - INFO - Processing hypothesis_agent agent
2024-12-24 16:23:04,252 - core.Logger - INFO - State: {'messages': [HumanMessage(content='\nfile_path:hotels.csv\nUse machine learning to perform data analysis and write complete graphical reports\n', additional_kwargs={}, response_metadata={})], 'process': 'Initial process content', 'hypothesis': 'Initial hypothesis content', 'process_decision': 'Initial process_decision content', 'visualization_state': 'Initial visualization_state content', 'search_state': 'Initial search_state content', 'code_state': 'Initial code_state content', 'report_section': 'Initial report_section content', 'quality_review': 'Initial quality_review content', 'needs_revision': False}
2024-12-24 16:23:04,259 - core.Logger - ERROR - Error while executing hypothesis_agent node: "Input to ChatPromptTemplate is missing variables {'needs_revision', 'report_section', 'hypothesis', 'process', 'visualization_state', 'quality_review', 'code_state', 'process_decision', 'search_state', 'agent_scratchpad'}.  Expected: ['agent_scratchpad', 'code_state', 'hypothesis', 'messages', 'needs_revision', 'process', 'process_decision', 'quality_review', 'report_section', 'search_state', 'visualization_state'] Received: ['messages', 'is_last_step', 'remaining_steps']\nNote: if you intended {needs_revision} to be part of the string and not a variable, please escape it with double curly braces like: '{{needs_revision}}'.\nFor troubleshooting, visit: https://python.langchain.com/docs/troubleshooting/errors/INVALID_PROMPT_INPUT "

As seen I have printed the state in the log timed 2024-12-24 16:23:04,252 and it has all the input variable. This logging logic is present before the agent.invoke call in the agent_node method.

I have turned the debugging and verbose on in create_react_agent and AgentExecutor.from_agent_and_tools respectively with the following output:

> Entering new AgentExecutor chain...
[-1:checkpoint] State at the end of step -1:
{'messages': []}
[0:tasks] Starting 1 task for step 0:
- __start__ -> {'code_state': 'Initial code_state content',
 'hypothesis': 'Initial hypothesis content',
 'intermediate_steps': [],
 'messages': [HumanMessage(content='\nfile_path:hotels.csv\nUse machine learning to perform data analysis and write complete graphical reports\n', additional_kwargs={}, response_metadata={})],
 'needs_revision': False,
 'process': 'Initial process content',
 'process_decision': 'Initial process_decision content',
 'quality_review': 'Initial quality_review content',
 'report_section': 'Initial report_section content',
 'search_state': 'Initial search_state content',
 'visualization_state': 'Initial visualization_state content'}
[0:writes] Finished step 0 with writes to 1 channel:
- messages -> [HumanMessage(content='\nfile_path:hotels.csv\nUse machine learning to perform data analysis and write complete graphical reports\n', additional_kwargs={}, response_metadata={})]
[0:checkpoint] State at the end of step 0:
{'messages': [HumanMessage(content='\nfile_path:hotels.csv\nUse machine learning to perform data analysis and write complete graphical reports\n', additional_kwargs={}, response_metadata={}, id='b701ca0a-287a-4a8b-8ed7-9381cc2cedaa')]}
[1:tasks] Starting 1 task for step 1:
- agent -> {'is_last_step': False,
 'messages': [HumanMessage(content='\nfile_path:hotels.csv\nUse machine learning to perform data analysis and write complete graphical reports\n', additional_kwargs={}, response_metadata={}, id='b701ca0a-287a-4a8b-8ed7-9381cc2cedaa')],
 'remaining_steps': 29}

As observed between [0:tasks] Starting 1 task for step 0: and [0:writes] Finished step 0 with writes to 1 channel: only the message key is retained

Can anyone help me understand what is happening and can I tackle the issue

[AkashBais]

@dosu can you explain and give a resolution to the issue