create_agent with tools fail as final LLM call does not happen and it errors out with UnboundLocalError

[hoppiesbunny]

Checked other resources

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).
• This is not related to the langchain-community package.
• I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example).
• I posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Example Code

Example code that I have when used to call AzureChatOpenAI

    agent = create_agent(
        model=self._client,
        tools=[search_web],
        system_prompt=system_prompt,
        state_schema=state_typed_dict,
        response_format=response_model,
    )
    raw_llm_result = agent.invoke(
        input={"messages": [{"role": "user", "content": user_prompt}]}
    )

    # structured output is a json string in result['messages'][-1].content
    structured_output = json.loads(raw_llm_result["messages"][-1].content)

Error Message and Stack Trace (if applicable)

UnboundLocalError("cannot access local variable 'last_ai_index' where it is not associated with a value")

This is the function inside LangChain that is erroring out repeatedly:

def _fetch_last_ai_and_tool_messages(
    messages: list[AnyMessage],
) -> tuple[AIMessage, list[ToolMessage]]:
    last_ai_index: int
    last_ai_message: AIMessage

    for i in range(len(messages) - 1, -1, -1):
        if isinstance(messages[i], AIMessage):
            last_ai_index = i
            last_ai_message = cast("AIMessage", messages[i])
            break

    tool_messages = [m for m in messages[last_ai_index + 1 :] if isinstance(m, ToolMessage)]
    return last_ai_message, tool_messages

Description

The moment tools are included, it will error out after the tool call, even though the tools return valid results.
This only happens after I've upgraded to Langchain v1.0. Things work fine if no tools are included. The nature of the tools do not matter.

You can see from my Langsmith trace that the 2nd LLM call after all the tools never happen.

You can see that the tool does return valid results. I've redacted the information for privacy reasons.

Before the upgrade to Langchain v1.0, you can see that there will always be final LLM call after tool calls. (Refer to image below). There was no change in my business logic when I upgraded to Langchain v1.0.

I dug deeper into the Langchain source code and saw in the docstring of the create_agent function that the second LLM call should be a deterministic one:

    The agent node calls the language model with the messages list (after applying
    the system prompt). If the resulting `AIMessage` contains `tool_calls`, the graph
    will then call the tools. The tools node executes the tools and adds the responses
    to the messages list as `ToolMessage` objects. The agent node then calls the
    language model again. The process repeats until no more `tool_calls` are
    present in the response. The agent then returns the full list of messages.

Do let me know if more information is needed!

System Info

python = "^3.11"
langgraph = ">=0.2.6"
langchain-openai = "^1.0.1"
langchain-anthropic = "^1.0.0"
langchain-google-vertexai = "^3.0.1"
langchain = "^1.0.2"
langchain-fireworks = "^1.0.0"

[Sinha-Ujjawal]

Can you please share a reproducible example? The below example you had shared has references to self._client what is that?

    agent = create_agent(
        model=self._client,
        tools=[search_web],
        system_prompt=system_prompt,
        state_schema=state_typed_dict,
        response_format=response_model,
    )
    raw_llm_result = agent.invoke(
        input={"messages": [{"role": "user", "content": user_prompt}]}
    )

    # structured output is a json string in result['messages'][-1].content
    structured_output = json.loads(raw_llm_result["messages"][-1].content)

[hoppiesbunny]

@Sinha-Ujjawal this is a reproducible example:

This is the search tool that I wanted to use:

import os
from typing import List, Optional
from langchain_core.tools import BaseTool
from langchain_tavily import TavilySearch

def get_tavily_search_tool(
    max_results: int = 5,
    search_depth: str = "advanced",
    include_domains: Optional[List[str]] = None,
    exclude_domains: Optional[List[str]] = None,
) -> BaseTool:

    api_key = os.getenv("TAVILY_API_KEY")

    tool = TavilySearch(
        max_results=max_results,
        search_depth=search_depth,
        include_domains=include_domains,
        exclude_domains=exclude_domains,
        api_key=api_key,
    )

    return tool

Then, inside my agent function, I initialised my tool, and passed it to the agent:

  search_web = get_tavily_search_tool()

  agent = create_agent(
      model='gpt-4.1',
      tools=[search_web],
      system_prompt=system_prompt,
      state_schema=state_typed_dict,
      response_format=response_model,
  )
  raw_llm_result = agent.invoke(
      input={"messages": [{"role": "user", "content": user_prompt}]}
  )

  # structured output is a json string in result['messages'][-1].content
  structured_output = json.loads(raw_llm_result["messages"][-1].content)

I was never able to get the raw_llm_result because it will error out with UnboundLocalError. You can see in my provided screenshots from Langsmith that it is missing a final LLM call after all the tool calling. This only started to happen after I upgraded to Langsmith v1.0. I wasn't experiencing this before the upgrade (refer to screenshot)

Thanks so much for your help!! Let me know if anything is unclear!

[Sinha-Ujjawal]

@hoppiesbunny Can you also tell me what is state_typed_dict and response_model, because I tried below example with Ollama model, and it worked fine for me.

import os
from typing import List, Optional

from langchain.agents import create_agent
from langchain_core.tools import BaseTool
from langchain_ollama import ChatOllama
from langchain_tavily import TavilySearch
from load_dotenv import load_dotenv

load_dotenv()


def get_tavily_search_tool(
    max_results: int = 5,
    search_depth: str = "advanced",
    include_domains: Optional[List[str]] = None,
    exclude_domains: Optional[List[str]] = None,
) -> BaseTool:
    api_key = os.getenv("TAVILY_API_KEY")

    tool = TavilySearch(
        max_results=max_results,
        search_depth=search_depth,
        include_domains=include_domains,
        exclude_domains=exclude_domains,
        api_key=api_key,
    )

    return tool


search_web = get_tavily_search_tool()

system_prompt = ""
user_prompt = ""

llm = ChatOllama(model="qwen3:4b")

agent = create_agent(
    # model="gpt-4.1",
    model=llm,
    tools=[search_web],
    system_prompt=system_prompt,
    # state_schema=state_typed_dict,  # COMMENTED THIS
    # response_format=response_model, # COMMENTED THIS
)
raw_llm_result = agent.invoke(
    input={"messages": [{"role": "user", "content": user_prompt}]}
)

print(raw_llm_result)

# structured output is a json string in result['messages'][-1].content
# structured_output = json.loads(raw_llm_result["messages"][-1].content)

[Sinha-Ujjawal]

BTW this is my pyproject.toml:

[project]
name = "temp"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.11"
dependencies = [
    "langchain>=1.0",
    "langchain-ollama>=1.0.0",
    "langchain-openai>=1.0.1",
    "langchain-tavily>=0.2.12",
    "load-dotenv>=0.1.0",
]

[Sinha-Ujjawal]

Your code is definitely failing in _fetch_last_ai_and_tool_messages function, and that can happen if len(messages) == 0 or there are not AIMessage instances in messages. But I believe the messages will always have at least one AIMessage for the system_prompt. The actual implementation should handle this properly, but I am not sure what exactly went wrong in your case.

I am definitely seeing linter errors for _fetch_last_ai_and_tool_messages, but really want to understand what went wrong in your case.