Agent can't stop on function calls and keeps iterating in an infinite loop on function tools

[SalmaAmrAhmed]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

import json
import os
from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain.agents import create_openai_functions_agent, AgentExecutor
from langchain_core.tools import tool
from langchain_core.messages import HumanMessage, AIMessage


# Define tools
@tool
def multiply(first_int: int, second_int: int) -> int:
    """Multiply two integers together."""
    return first_int * second_int


@tool
def add(first_int: int, second_int: int) -> int:
    """Add two integers."""
    return first_int + second_int


@tool
def exponentiate(base: int, exponent: int) -> int:
    """Exponentiate the base to the exponent power."""
    return base ** exponent


class ConversationalExecutor:
    def __init__(self):

        self.llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0)
        self.tools = [multiply, add, exponentiate]

        self.prompt = ChatPromptTemplate.from_messages([
            MessagesPlaceholder(variable_name="messages"),
            MessagesPlaceholder(variable_name="intermediate_steps"),
            MessagesPlaceholder(variable_name="agent_scratchpad")
        ])

        self.prompt.messages = [AIMessage("You are a very powerful assistant, you can use a set of tools and return the final result to the user")]

        self.agent = create_openai_functions_agent(self.llm, self.tools, self.prompt)

        self.agent_executor = AgentExecutor(
            agent=self.agent,
            tools=self.tools,
            verbose=True,
            max_iterations=3,
            return_intermediate_steps=True,
            early_stopping_method="force"
        )

    def handle_conversation(self, user_input):
        self.prompt.messages.append(HumanMessage(user_input))
        output = self.agent_executor.invoke({
            "messages": self.prompt.messages
        })
        print(output['intermediate_steps'])
        return output['output']


if __name__ == '__main__':
    intro = 'How can I help you?'
    print(f'Agent: {intro}')
    user_input = input()
    agent = ConversationalExecutor()
    agent.prompt.messages.append(AIMessage(intro))
    while user_input.strip() != '':
        results = agent.handle_conversation(user_input)
        print('Agent:', results)
        user_input = input()

### Description

I am trying to build an abstract class to handle conversational flow in an app, testing on simple input like 2+3 and it gets stuck in an infinite loop, unable to parse the output. I add `max_iterations` to stop, however it shouldn't get stuck in such a simple request. Can you help me figure the issue?

Below you can see my output:

Agent: How can I help you?
2+3

Entering new AgentExecutor chain...

Invoking: add with {'first_int': 2, 'second_int': 3}

5
Invoking: add with {'first_int': 2, 'second_int': 3}

5
Invoking: add with {'first_int': 2, 'second_int': 3}

5

Finished chain.
[(AgentActionMessageLog(tool='add', tool_input={'first_int': 2, 'second_int': 3}, log="\nInvoking: add with {'first_int': 2, 'second_int': 3}\n\n\n", message_log=[AIMessageChunk(content='', additional_kwargs={'function_call': {'arguments': '{"first_int":2,"second_int":3}', 'name': 'add'}})]), 5), (AgentActionMessageLog(tool='add', tool_input={'first_int': 2, 'second_int': 3}, log="\nInvoking: add with {'first_int': 2, 'second_int': 3}\n\n\n", message_log=[AIMessageChunk(content='', additional_kwargs={'function_call': {'arguments': '{"first_int":2,"second_int":3}', 'name': 'add'}})]), 5), (AgentActionMessageLog(tool='add', tool_input={'first_int': 2, 'second_int': 3}, log="\nInvoking: add with {'first_int': 2, 'second_int': 3}\n\n\n", message_log=[AIMessageChunk(content='', additional_kwargs={'function_call': {'arguments': '{"first_int":2,"second_int":3}', 'name': 'add'}})]), 5)]
Agent: Agent stopped due to iteration limit or time limit.

### System Info

System Information
------------------
> OS:  Darwin
> OS Version:  Darwin Kernel Version 22.3.0: Mon Jan 30 20:38:37 PST 2023; root:xnu-8792.81.3~2/RELEASE_ARM64_T6000
> Python Version:  3.10.13 | packaged by conda-forge | (main, Oct 26 2023, 18:09:17) [Clang 16.0.6 ]

Package Information
-------------------
> langchain_core: 0.1.26
> langchain: 0.1.8
> langchain_community: 0.0.21
> langsmith: 0.1.5
> langchain_openai: 0.0.7
> langchainhub: 0.1.14

Packages not installed (Not Necessarily a Problem)
--------------------------------------------------
The following packages were not found:

> langgraph
> langserve

[dosubot[bot]]

🤖

Hey @SalmaAmrAhmed, great to see you back! Hope you're doing well.

Based on the code you've provided, it seems like the AgentExecutor is repeatedly invoking the add function because it's not able to recognize that the function has already been successfully executed and the result has been obtained. This is likely due to the way the AgentExecutor is handling the output from the add function.

In your ConversationalExecutor class, you're using the invoke method of the AgentExecutor to execute the agent with the given messages. The invoke method is designed to execute the agent until a stopping condition is met, which in your case is set to force and a maximum of 3 iterations. However, it seems like the stopping condition is not being met, causing the AgentExecutor to keep invoking the add function.

One possible reason for this could be that the AgentExecutor is not able to correctly interpret the output from the add function. The AgentExecutor uses the output from the function to determine the next action to take. If it's not able to correctly interpret the output, it might keep invoking the function in an attempt to get a valid output.

To resolve this issue, you might need to adjust the way the AgentExecutor handles the output from the function. This could involve modifying the invoke method or the way the output from the function is parsed.

You can refer to the following resources for more information on how the AgentExecutor works and how to customize its behavior:

• AgentExecutor source code
• LangChain documentation on Agents

I hope this helps! If you have any other questions or need further clarification, feel free to ask.

Sources

• Issue : Agent Executor stops due to iteration limit or time limit.
• Agent Executor stuck in loop
• libs/core/langchain_core/callbacks/manager.py
• libs/langchain/langchain/agents/agent_toolkits/openapi/base.py

About Dosu

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[eyurtsev]

Follow instructions here -- https://python.langchain.com/docs/modules/agents/how_to/custom_agent#create-the-agent

agent definition doesn't look right

[nhathoang0110]

@nikhitaKanoj did u get the solution ?

[JiitBhatt]

@nhathoang0110 did you get the solution?

[Harrison-Li]

@JiitBhatt did you get the solution?

[weifxn]

@Harrison-Li did you get the solution?

[Harrison-Li]

@Harrison-Li did you get the solution?

My problem is quite easy which might not be the problem in your case, my tool is using hard code to output the image, while this image output can not be sent to the LLM supervisor agent. My suggestion would be to check your tool output and whether your LLM received the answer from the tool.