Langchain agent loops over the same tool producing the same result.

[klychliiev]

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

prompt template:

CHAT_PROMPT_TEMPLATE = """
Respond to the human as helpfully and accurately as possible. You have access to the following tools:
{tools}
Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).
Valid "action" values: "Final Answer" or {tool_names}
Provide only ONE action per $JSON_BLOB, as shown:

{{
"action": $TOOL_NAME,
"action_input": $INPUT
}}

Follow this format:
Question: input question to answer
Thought: consider previous and subsequent steps
Action:

$JSON_BLOB

Observation:
the result of the action
... (Proceed to Final Answer once the tool is executed)
Thought: I know what to respond
Action:

{{
"action": "Final Answer",
"action_input": "Final response to human"
}}
Once the agent is called, run the chosen tool only once. You shouldn't loop over the same tool.
If the tool executed successfully, write a positive response to the user empohasizing success of the action and stop the agent.
Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB Observation
Consider the previous chat history:
{chat_history}
New input: {input}
{agent_scratchpad}
"""

function to create agent:

    def create_agent(self, selected_model: str, memory: ConversationBufferMemory, chat_name: str) -> AgentExecutor:
        """
        Function to generate the chain using prompt template
        and chat client (Claude or GPT).

        Args:
            selected_model (str): the LLM chosen by user:
            Claude 3 Opus by Anthropic or GPT-3.5 Turbo by OpenAI

        Returns:
            chain (RunnableSequence): chain incorporating prompt template
            and chosen language model
        """
        if selected_model == "GPT-3.5":
            chat_client = ChatOpenAI(api_key=OPENAI_API_KEY, model="gpt-3.5-turbo-1106")
        else:
            chat_client = ChatAnthropic(
                anthropic_api_key=ANTHROPIC_API_KEY, model_name="claude-3-opus-20240229"
            )

        tools = [TavilySearchResults(chat_name=chat_name),
                 DatasetSummarizer(chat_name=chat_name),
                 EDAAutomator(chat_name=chat_name),
                 DataVisuzalizer(chat_name=chat_name),
                 VisualizationExplainer(chat_name=chat_name),
                 VisualizationEditor(chat_name=chat_name),
                 VisualizationRecommender(chat_name=chat_name),
                 VisualizationRepair(chat_name=chat_name),
                 TextToSQL(chat_name=chat_name),
                 MathSolver(chat_name=chat_name),
                 BinderExecutor(chat_name=chat_name)]

        agent = create_structured_chat_agent(
            chat_client,
            tools,
            self.prompt
        )

        return AgentExecutor(
            agent=agent,
            tools=tools,
            verbose=True,
            memory=memory,
            handle_parsing_errors="If the tool executed successfully, stop the agent and write positive 'Final Answer' emphasizing success of the action",
        )

Description

Agent loops over the same tool producing the same result.
Console output:

Entering new AgentExecutor chain...
{
"action": "visualization_recommender",
"action_input": {
"recommend_data": {
"number_of_recommendations": 1,
"visualization": "d6b37fe7-61a9-4390-bccc-7b6c5e8cdc56"
}
}
}
Recommended chart generated successfully{
"action": "visualization_recommender",
"action_input": {
"recommend_data": {
"number_of_recommendations": 1,
"visualization": "d6b37fe7-61a9-4390-bccc-7b6c5e8cdc56"
}
}
}
Recommended chart generated successfully```
{
"action": "visualization_recommender",
"action_input": {
"recommend_data": {
"number_of_recommendations": 1,
"visualization": "d6b37fe7-61a9-4390-bccc-7b6c5e8cdc56"
}
}
}
Recommended chart generated successfully{
"action": "visualization_recommender",
"action_input": {
"recommend_data": {
"number_of_recommendations": 1,
"visualization": "d6b37fe7-61a9-4390-bccc-7b6c5e8cdc56"
}
}
}
Recommended chart generated successfully

System Info

langchain==0.1.9
langchain-community==0.0.24
langchain-core==0.1.26
langchain-openai==0.0.7
langchain-anthropic==0.1.4

[dosubot[bot]]

To address the issue where your AgentExecutor is looping over the same tool and producing the same result, it's important to investigate a few potential causes:

1. Check the Agent's Planning Logic: The behavior you're observing might be due to the agent's planning logic repeatedly selecting the same tool for execution. Review the implementation of the plan or aplan method in your agent class to ensure it's designed to select different tools based on the current state and inputs.

2. Review Stopping Conditions: If the stopping conditions for the agent's execution loop are not being met, it could lead to repetitive execution of the same tool. Verify the configuration of parameters like max_iterations and max_execution_time to ensure they are set to prevent endless loops.

3. Update LangChain: You're currently using LangChain version 0.1.9. Since there have been updates to LangChain, consider upgrading to the latest version to benefit from any fixes or improvements that might address your issue.

By focusing on these areas, you should be able to diagnose and resolve the looping behavior you're experiencing with the AgentExecutor.

Sources

• Agent Executor stuck in loop
• Don't try tools that have already been invalidated in the current run
• libs/langchain/langchain/agents/tool_calling_agent/base.py
• libs/langchain/langchain/agents/agent.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.

[Joshua-Shepherd]

This is also happening to me. Albeit intermittently.
I can reliably reproduce the issue when attaching a CallbackHandler to the AgentExecutor.
super fucking annoying. I'll try digging into the AgentExecutor code and see if I can find anything helpful to share

[klychliiev]

will be grateful for any hints!

[BrunoBrasolin]

Do you guys figure it out on how to stop the loop? Looks like the LLM doesnt understand the observation

This is my prompt and the agent is making a loop on the same tool forever
https://smith.langchain.com/prompts/dalme/df541d35?organizationId=29dafe83-fe58-5faf-90dd-59237b3d92bd

[BrunoBrasolin]

Just found the answer, my prompt was a ChatPromptTemplate and not a PromptTemplate

Works:

Loop:

[alonme]

Hey @BrunoBrasolin!

Can you please explain why this change solves the issue?

[BrunoBrasolin]

Hi, @alonme!

On my langchain hub I was creating an PromptTemplate instead of a ChatPromptTemplate. When I change to ChatPromptTemplate the problem was solved!

[alonme]

I Solved this issue by changing the prompt while still using ChatPromptTemplate

Changed from this

    prompt = ChatPromptTemplate.from_messages(
        [
            ("system", get_base_prompt()),
            MessagesPlaceholder(variable_name="history", optional=True),
            (
                "human",
                """

            {input}


            {agent_scratchpad}""",
            ),
        ]
    )

To this

    prompt = ChatPromptTemplate.from_messages(
        [
            ("system", get_base_prompt()),
            MessagesPlaceholder(variable_name="history", optional=True),
            ("human","{input}"),
            MessagesPlaceholder(variable_name="agent_scratchpad", optional=True),
        ]
    )