Is it possible to just return the response from a tool directly?

[hodgesz]

I am working on a PoC and modified this example I found online - https://github.com/christianrice/ai-demos/blob/96050c43f06cf43ff059458c8d19d49dad49da3b/2024_02_09%20-%20LangGraph%20Parallel%20Tool%20Calling/03_LangGraph_LLM_with_tools.ipynb

I have it working where it correctly calls the tools and returns the appropriate response, but there is an extra LLM call/step at the end by the agent I am trying to remove.

Here is the graph definition:

Here is the 'should_continue' function:

You can see in the following LangSmith screen shot the 3 calls (agent > action (tool) > agent)

Is it possible to just return the answer directly from action since no new information is added by the final agent/llm call?

[hodgesz]

For a little more clarity here is a pic of the graph.

Additionally, I set the tool parameter 'return_direct' to True but still has the extra final call through the agent I am trying to avoid or remove.

question_answer = StructuredTool.from_function(
func=ask_question,
name="Ask_Question",
description="Answer the question from the student.",
args_schema=AskQuestion,
return_direct=True,
)

[hodgesz]

I figured out my issue and a couple different ways to solve by adjusting the structure of the graph. I could have a conditional edge from action to end, which works pretty well. Appreciate the help!

[chaliy]

Could you pls post, your solution?

[hinthornw]

Awesome! Ya return_direct doesn't really mean anything in langgraph, that option waas for LangChain's agent executor

[chaliy]

Hello @hinthornw , any plans to add this functionality directly to create_react_agent? It seems like quite logical functionality.

Here is related issue in langchain repo - langchain-ai/langchain#22441