Agent output is just one line.

[invalidexplorer]

Hi! I am starting off with langgraph and I am trying to build a simple feedback multiagent loop.

`import getpass
import os


def _set_if_undefined(var: str):
    if not os.environ.get(var):
        os.environ[var] = getpass.getpass(f"Please provide your {var}")


_set_if_undefined("LANGCHAIN_API_KEY")

# Optional, add tracing in LangSmith
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_PROJECT"] = "Multi-agent Collaboration 3"

from langchain_aws import ChatBedrockConverse


llm = ChatBedrockConverse(model_id="anthropic.claude-3-sonnet-20240229-v1:0")


class GraphState(TypedDict):

    keys: Dict[str, any]

from typing import Annotated

from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_core.tools import tool
from langchain_experimental.utilities import PythonREPL
repl = PythonREPL()


@tool
def python_repl(
    code: Annotated[str, "The python code to execute to parse unstructured multiple worksheet excel."],
):
    """Use this to execute python code. If you want to see the output of a value,
    you should print it out with `print(...)`. This is visible to the user."""
    try:
        result = repl.run(code)
    except BaseException as e:
        return f"Failed to execute. Error: {repr(e)}"
    result_str = f"Successfully executed:\n```python\n{code}\n```\nStdout: {result}"
    return (
        result_str + "\n\nIf you have completed all tasks, respond with FINAL ANSWER."
    )

from langchain_experimental.tools import PythonREPLTool
from langchain_core.prompts import (
    ChatPromptTemplate,
    MessagesPlaceholder,
    SystemMessagePromptTemplate,
        HumanMessagePromptTemplate,

)
def create_agent(system_message: str):

    python_repl_tool = python_repl
    tools = [python_repl_tool]
    # use Sonnet or Haiku models not claude instant or 2.1 if asked for a gen AI script
    human_template = "{input}"

    prompt1 = [
        SystemMessagePromptTemplate.from_template(system_message),
        HumanMessagePromptTemplate.from_template(input_variables=["input"], template=human_template),
        MessagesPlaceholder(variable_name="question", optional=True),
        MessagesPlaceholder(variable_name="agent_scratchpad", optional=True),
    ]

    prompt = ChatPromptTemplate.from_messages(prompt1)
    return prompt | llm.bind_tools(tools) | StrOutputParser()



summarizer = create_agent("""<instructions> You are a programmer that has to understand the structure of a semi structure excel using python. Your input is only the path {path} to the excel. You can only use pandas. Give summary of each worksheet, and each table in the worksheet. Remember each worksheet can have multiple tables. Your job is to give information about the only relevant tables (will and should be well structured and will start after and only after row 3)
   + special request: {request} </instructions>""")
coder = create_agent('''You have the access to license statement file at {path} and summary of what the document looks like which is {summary}. The file is in a semi-structured format,
generate the code to parse License Summary sheet in the file. Here is the overall procedure:
1. Decide the exact worksheet name. If can't find the worksheet, read the sheet_names firstly and decide the which worksheet to read.
2. Read the worksheet with pandas dataframe
3. Iterate each row and enumerate each column. Find a structured table, NOT random unstructured filler words like description. Find columns "related" to "License count" and "product name". Find all rows related with columns. *Note* it is not neccessary that the row names will be license count or product name, use your own intellegence to figure that out, first read the columns names and then see which ones are relevant.
4. If there are both count and product name in a row, remember the row index as header row. remember both indexes, break
4. If contains only one of non of them, reset header row index and both column indexes.
5. Starting from header row onwards, continue to read the corresponding name of the product and license count columns and so on, skip all the empty or na values, print valid ones.
6. Convert the data into JSON format. With column names being the key and rows being the value.
Once you see the generated code, execute the generated python code with the tool. Remeber python code has all the access to local files.

general approach of reading Excel data into a pandas DataFrame, locating the relevant rows/columns, extracting the needed data, and converting to JSON format.
                     
When you are done add "___FINISHED___" at the end of the response. 
If you are not able to do this, ask summarizer about what information you need it to find. (in your response make it like an effective prompt)''')

def excel_summary_node(state):
    print("---summary excel---")
    state_dict = state["keys"]
    path = state_dict["path"]
    userInput = state_dict["input"]
    if(state_dict["request"]):
        request = state_dict["request"]
    else:
        request = ''
    summary = summarizer.invoke({"path" : path, 
                                  "request" : request, "input" : userInput
                                     })
    state_dict["summary"] = summary
    print("summary excel done"
          )
    print(summary)

    return {"keys": state_dict}

def coder_node(state):

    print("---coder---")
    state_dict = state["keys"]
    path = state_dict["path"]
    summary = state_dict["summary"]
    userInput = state_dict["input"]
    code = coder.invoke({"path" : path, "summary" : summary, "input" : userInput })

    if "FINAL ANSWER" in code:
        state_dict["END"] = True

    state_dict["code"] = code
    print("Coder done")
    print(code)

    return {"keys": state_dict}

from typing import Literal


def router(state) -> Literal["call_tool", "__end__", "continue"]:
    state_dict = state["keys"]
    if state_dict.get("END", False):
        # Any agent decided the work is done
        return "__end__"
    return "continue"

workflow = StateGraph(GraphState)
from langgraph.graph import END, StateGraph, START

workflow.add_node("coder", coder_node)
workflow.add_node("summarizer", excel_summary_node)

workflow.add_edge(START, "summarizer")
workflow.add_edge("summarizer", "coder")
workflow.add_conditional_edges(
    "coder",
    router,
    {"continue": "summarizer", "__end__": END},
)

app = workflow.compile()

inputs = { "keys":{"path":'LS.xlsx', "request":"",  "input": "convert it into json"}}
graph_out = app.invoke(inputs, {"recursion_limit": 8})
print(graph_out["keys"]["code"])`
However the outputs are just one liners, like 

`
---summary excel---
summary excel done
Here is how we can use Python and pandas to summarize the structure of the Excel file "LS.xlsx":
---coder---
Coder done
Here is some Python code to parse the "License Summary" sheet in the LS.xlsx file and convert the relevant data to JSON:
---summary excel---
summary excel done
Here is how we can parse the Excel file 'LS.xlsx' using Python and pandas:
---coder---
Coder done
Here is the Python code to parse the 'License Summary' sheet from the 'LS.xlsx' file and convert the data into JSON format:
---summary excel---
summary excel done
Here is how we can parse the semi-structured Excel file using Python and pandas:
---coder---
Coder done
Here is how we can parse the semi-structured Excel file into JSON format using Python and pandas:
---summary excel---
summary excel done

---coder---
Coder done

Here is the Python code to parse the License Summary sheet in the provided Excel file LS.xlsx and convert it to JSON format: `

I feel my approach is wrong, how can I improve my code to be better and more robust? and How can I add a scratch pad into my agents?

[hwchase17]

this is most likely related to prompting. can you share a langsmith trace? it will make it easier to debug

but without more info I would guess this is pretty related to prompting and you just need to do some prompt engineering

[invalidexplorer]

Hi! @hwchase17 Here is the langsmith trace:
https://smith.langchain.com/public/4ae3f208-2297-4032-a0e0-d43bdd47b686/r

It is weird because it is able to run the program in REPL but it is not able to communicate that program with the other agent.