Langgraph did not call tools

[swarnitwayal]

Checked other resources

• This is a bug, not a usage question. For questions, please use GitHub Discussions.
• I added a clear and detailed title that summarizes the issue.
• I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example).
• I included a self-contained, minimal example that demonstrates the issue INCLUDING all the relevant imports. The code run AS IS to reproduce the issue.

Example Code

from datetime import datetime
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent
from config import BASE_URL, API_TOKEN, APPLICATION_TYPE_HEADER

class CustomLLM(ChatOpenAI):
    def __init__(self, api_token: str, base_url: str, application_type: str, **kwargs):
        default_headers = {
            "applicationType": application_type,  # This is the header you want to include
        }
        super().__init__( model='gpt-4o', openai_api_base=base_url, openai_api_key=api_token, default_headers=default_headers, **kwargs)

def check_weather(location: str, at_time: datetime | None = None) -> str:
	'''Return the weather forecast for the specified location.'''
	return f"It's always sunny in {location}"

tools = [check_weather]
model = CustomLLM(base_url=BASE_URL, api_token=API_TOKEN, application_type=APPLICATION_TYPE_HEADER)

graph = create_react_agent(model, tools)
inputs = {"messages": [("user", "what is the weather in sf")]}

for s in graph.stream(inputs, stream_mode="values"):
	message = s["messages"][-1]
	if isinstance(message, tuple):
		print(message)
	else:
		message.pretty_print()

Error Message and Stack Trace (if applicable)

(.venv) PS C:\Repositories\Langgraph-latest> & c:/Repositories/Langgraph-latest/.venv/Scripts/python.exe c:/Repositories/Langgraph-latest/weather_botv3.py
================================ Human Message =================================

what is the weather in sf
================================== Ai Message ==================================

I'm sorry, I do not have real-time information on current weather conditions. I recommend checking a weather website or app for the most up-to-date weather forecast for San Francisco.

Description

I see that LLM doesn't make any tool calls with above code example.

System Info

System Windows:

langchain==0.3.20
langchain-core==0.3.44
langchain-openai==0.3.8
langchain-text-splitters==0.3.6
langgraph==0.3.7
langgraph-checkpoint==2.0.18
langgraph-prebuilt==0.1.2
langgraph-sdk==0.1.55
langsmith==0.3.13

Python 3.12.9

[hinthornw]

Hiya - your model isn't using its tools effectively. If i replace with a regular openai, gemini, or anthropic model here it works as expected.

Does your model gateway (or whatever you're connecting to there) support the tools arguments in the API?

(In other words, this doesn't appear to be an issue with langgraph)

[hinthornw]

Will convert this to a discussion, as this is not a langgraph bug.

[swarnitwayal]

Hi @hinthornw

The same example when tried with LangChain instead of LangGraph work perfectly fine with tool calls on same LLM.

from langchain.agents import initialize_agent, AgentType
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from config import BASE_URL, API_TOKEN, APPLICATION_TYPE_HEADER


# Custom LLM with API token 
class CustomLLM(ChatOpenAI):
    def __init__(self, api_token: str, base_url: str, application_type: str, **kwargs):
        default_headers = {
            "applicationType": application_type,  # This is the header you want to include
        }
        super().__init__(openai_api_base=base_url, openai_api_key=api_token, default_headers=default_headers, **kwargs)

# Tool to get weather
@tool
def get_weather(location: str):
    """Call to get the weather from a specific location."""
    print(f"Received location: {location}")  # Debug print
    if any([city in location.lower() for city in ["sf", "san francisco"]]):
        return "It's sunny in San Francisco."
    else:
        return f"I am not sure what the weather is in {location}"

# Instantiate Custom LLM
custom_llm = CustomLLM(api_token=API_TOKEN, base_url=BASE_URL, application_type=APPLICATION_TYPE_HEADER)

# Register tools
tools = [get_weather]

# Initialize the Agent with AgentExecutor
agent = initialize_agent(
    tools, 
    custom_llm, 
    agent_type=AgentType.ZERO_SHOT_REACT_DESCRIPTION,  # This agent type is for simple reactive agents
    verbose=True  # This will log the internal reasoning of the agent
)



# Function to print the message flow
def print_agent_output(inputs):
    output = agent.run(inputs)
    print(f"Agent Response: {output}")

# Define the input message (the user's request)
inputs = "What is the weather in SF?"

# Display the message flow
print_agent_output(inputs)

Output:

> Entering new AgentExecutor chain...

I should use the get_weather tool to find out the weather in San Francisco.
Action: get_weather
Action Input: "SF"Received location: SF

Observation: It's sunny in San Francisco.
Thought:
I now have the information I need to answer the question.
Final Answer: It's sunny in San Francisco.

> Finished chain.
Agent Response: It's sunny in San Francisco.

[raydrak64]

The current state of the code.
Using the tool serpapi and the already exampled check_weather to test tools.

'''
from langchain_ollama import ChatOllama
from langchain_core.messages import SystemMessage
from langchain_community.agent_toolkits.load_tools import load_tools
from langgraph.graph import StateGraph, MessagesState, START
from langgraph.graph.message import add_messages
from langgraph.checkpoint.sqlite import SqliteSaver
from langgraph.prebuilt import ToolNode, tools_condition

memory = SqliteSaver(memory_db)
tools = load_tools(["serpapi"])
if not tools:
print("\nCORE INIT: Failed to load tools")
else:
print("\nCORE INIT: Tools loaded successfully")
serpapi_tool = tools[0]
adax_tools = [serpapi_tool, check_weather]

workflow = StateGraph(State)

workflow.add_node("adax", self.llm_response)
workflow.add_node("tools", ToolNode(adax_tools)) # (tools=[tool])

workflow.add_edge(START, "adax")
workflow.add_conditional_edges("adax", tools_condition)
workflow.add_edge("tools", "adax")

self.graph = workflow.compile(checkpointer=memory)

self.llm = ChatOllama(xxx=xxx).bind_tools(adax_tools)

self.whatisaid = self.llm.invoke(messages)
'''
Messages:
'''
content='' additional_kwargs={} response_metadata={'model': 'qwen3:8b', 'created_at': '2025-05-15T18:55:13.9955115Z', 'done': True, 'done_reason': 'stop', 'total_duration': 5616805800, 'load_duration': 1180049100, 'prompt_eval_count': 804, 'prompt_eval_duration': 489325600, 'eval_count': 170, 'eval_duration': 3940525800, 'model_name': 'qwen3:8b'} id='run--e1a8aeac-e553-4cc7-95a4-4e2a0d63f987-0' tool_calls=[{'name': 'check_weather', 'args': {'location': 'Los Angeles'}, 'id': '1a156ed8-1f72-4e52-b28b-622cc4d9ee7d', 'type': 'tool_call'}] usage_metadata={'input_tokens': 804, 'output_tokens': 170, 'total_tokens': 974}
'''

It seems, with investigative testing and research, IMHO, tools just aren't executing under langgraph.

I went through the phase:
It's me, I am missing something small.

So, I believe I have tried every possible code combination. Using START, END, or neither and used
entry_node, prebuilt tools_condition or the custom working example commonly used: should_continue.
You can see my mermaid graph is correct, at least according to all the sources.

As for the models I have tried Mistral (and the type), which is supposed to be known to work
with ChatOllama and tools. They deny the tools exist. No matter my temperature or changing
the graph so tools is in a forced workflow.

Now, qwen2.5, qwen3 and the type (IE Cogito), recognize the tools. According to LangSmith,
the tools are being called. As per my example 'messages' you see the tool is being called.

However, both the 'messages' show a null output and langsmith says the output of the tool
is END with no tool tokens used. Which that may be the correct output but I would think I
would see the tools output with tokens being used...? This is true for both the prebuilt tool
serpapi and the custom check_weather. I put print statements in check_weather, it seems to
never execute.

Also to put on the table, serpapi's dashboard says there has never been a call made to the
service. Though all signs point to it being called.

So, one place I read, they do a NULL check on the response output of langgraph and if there
is a NULL reply with a tool call they do a second call: llm.invoke(messages).tool_call (or similiar)

I tried it, still got a NULL response. That logic didn't feel right anyways because isn't
that what the graphs workflow is to do, execute the tool if it's called? Then pass the repsonse
back to the agent to finalize the reply.

Which brings me to the last part that confuses me. If the tool did give NULL as a response,
wouldn't the agent still create a response saying the tool didn't help or something along
those lines? Give some kind of response, made up, hullicanted whatever.

Anywho, that was my 3 cents, sorry for the extra penny. :)

[raydrak64]

@swarnitwayal I figured out what I was doing wrong. Well, not really, I know what it was, just not why. I hope you figured your end out too. As for me...

I was needing my agent instance to be an object. So, I was taking all the lessons and examples out there that are functional and translating them into objects or classes. At first, I did it off the cuff and though the logic was there and correct, the agent could chat all day but, the tools just wouldn't fire. Since I didn't start from a working example, I didn't know what a correct trace looked like.

I took the prebuild example as-is and of course, it did work. I believe you said the one you supplied wasn't working, but it did. Then I slowly started to move it into a class. At one point, I started to get the same results using the prebuilt. The tool gets called then nothing. So, I would start again until I finally got an agent object.

I've been working with AI since the 80's and the whole point of me trying this was so I would stop reinventing the wheel. So, using the prebuilt was perfect for me. I can still add all the bells and whistles and my spin from my decades of research, and I don't have to work so hard keeping up.