Prevent an agent from answering anything after executing a certain tool

[yojimbo94]

Hi

I need to be able to execute a certain tool with an agent, but i don't want the agent to give any answer to the request.

All i want is :

• user talks in the chat
• agent detects a tool to use
• executes it
• doesn't do anything else (the tool triggers something on the front side and i don't want the agent to speak after tool's execution)

I can't use an empty "return direct" since i use a multi-action agent.

I don't want ALL of my tools to have that behavior, i wan't a classic agent usage with some tools that just don't return anything

Would love some help on this !
this is for langchain JS btw

[dosubot[bot]]

🤖

Hello @fgendre-u-iris!

I'm a friendly bot, Dosu, here to assist you while we wait for a human maintainer. I specialize in solving bugs, answering questions, and helping users become contributors. So, feel free to ask anything related to the LangChain repository. 😊 Let's make your LangChain experience smoother together!

Based on the current design of the LangChain framework, it seems that the agent's responses are tightly coupled with the execution of the tools. The agent's responses are constructed based on the intermediate steps of the tool execution, and there doesn't appear to be a built-in way to configure individual tools to not return any output after execution.

However, you might be able to achieve this by modifying the _construct_scratchpad method in the StructuredChatAgent class. You could add a condition to check if the current tool is one of the tools that should not return any output. If it is, you could return an empty string or None instead of the agent's scratchpad. Here's a rough example of what this could look like:

def _construct_scratchpad(
    self, intermediate_steps: List[Tuple[AgentAction, str]]
) -> str:
    agent_scratchpad = super()._construct_scratchpad(intermediate_steps)
    if not isinstance(agent_scratchpad, str):
        raise ValueError("agent_scratchpad should be of type string.")
    if agent_scratchpad:
        # Check if the current tool should not return any output
        if self.current_tool in tools_that_should_not_return_output:
            return ""
        else:
            return (
                f"This was your previous work "
                f"(but I haven't seen any of it! I only see what "
                f"you return as final answer):\n{agent_scratchpad}"
            )
    else:
        return agent_scratchpad

Please note that this is just a rough idea and might not work as expected without further modifications. Also, this would require you to maintain a list of tools that should not return any output (tools_that_should_not_return_output in the example), and you would need a way to access the current tool (self.current_tool in the example) from within the _construct_scratchpad method.

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

Sources

• libs/langchain/langchain/agents/structured_chat/base.py

---

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.

[yojimbo94]

If i do that, agent will still answer something (and the risk is that it could hallucinate or loop on the same tool)

[hwchase17]

just i understand, the steps you want are:

• user input
• agent decides potentially multiple tools to call
• some tools return things - in those cases, you want to continue in the normal agent loop
• some tools dont return things - if the agent calls only one tool, and it is this type of tool, you want to call the call and then stop the loop completely

what if it calls multiple tools and one of those is such a tool?

[yojimbo94]

these are exactly the steps i want.

Regarding your question, i don't mind stopping the loop in such a case.

Do you have any clues for this problem ?

[palvarezcordoba]

You could use astream_events method of AgentExecutor.
Doing so, you will get events. After the tool end event, you check what tool it was, and stop execution.