LangGraph Prebuilt Agents

Warnings

• gotcha LangGraph is a low-level framework, while `langgraph-prebuilt` offers higher-level abstractions. Mixing low-level graph construction (e.g., `StateGraph`, `add_node`) with prebuilt agents (like `create_react_agent`) without understanding LangGraph's state management can lead to unexpected behavior or difficult-to-debug issues. Stick to one paradigm or thoroughly understand how state is managed across both. [3, 5, 6]
  Fix: For complex customizations, consider extending or modifying the prebuilt agent's internal graph if the documentation provides a clear path, or build a custom graph from scratch using `langgraph.graph.StateGraph` for full control. For simpler cases, leverage the provided prebuilt agent parameters and hooks. Consult the official LangGraph documentation for advanced patterns. [5, 8, 9]
• gotcha Prebuilt agents often rely on external services (LLMs, search tools). Failing to set required API keys as environment variables (e.g., `OPENAI_API_KEY`, `TAVILY_API_KEY`) or passing incorrect values will cause agents to fail silently or with authentication errors. [1, 6, 13]
  Fix: Always verify that all necessary API keys are correctly set in your environment or explicitly passed to the respective LLM/tool constructors before running the agent. Test credentials independently if an agent is failing. Use a `.env` file and `python-dotenv` for local development. [1, 6]
• deprecated The LangGraph ecosystem, including `langgraph-prebuilt`, is actively developed. Older patterns for agents or tool integration, while functional, might be superseded by more efficient, flexible, or recommended approaches in newer releases. [7]
  Fix: Regularly consult the official LangChain/LangGraph documentation and quickstarts for the latest recommended practices and agent architectures. Pay attention to release notes for potential API changes or new high-level abstractions. [1, 7, 10]

Install

• pip install langgraph-prebuilt langchain-openai tavily-python Install core library and common dependencies

Imports

• create_react_agent

  from langgraph.prebuilt import create_react_agent

• ToolNode

  from langgraph.prebuilt import ToolNode

• tools_condition

  from langgraph.prebuilt import tools_condition

• MessagesState

  from langgraph.graph import MessagesState

  MessagesState is part of the core langgraph graph module, not prebuilt.

Quickstart

This quickstart demonstrates how to set up and use `create_react_agent` from `langgraph-prebuilt`. It initializes an OpenAI LLM, a Tavily search tool, and then creates an agent capable of using these tools. The agent is then invoked with a query, showcasing its ability to use tools for information retrieval and simple calculations. Ensure `OPENAI_API_KEY` and `TAVILY_API_KEY` are set in your environment for the example to run. [8, 13]

import os
from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults
from langgraph.prebuilt import create_react_agent
from typing import Annotated, Sequence, TypedDict
from langchain_core.messages import BaseMessage

# Set API keys (replace with actual keys or set as environment variables)
os.environ["OPENAI_API_KEY"] = os.environ.get("OPENAI_API_KEY", "YOUR_OPENAI_API_KEY")
os.environ["TAVILY_API_KEY"] = os.environ.get("TAVILY_API_KEY", "YOUR_TAVILY_API_KEY")

# Define the agent state (LangGraph requires a defined state)
class AgentState(TypedDict):
    messages: Annotated[Sequence[BaseMessage], lambda x, y: x + y]

# Initialize LLM and tools
llm = ChatOpenAI(model="gpt-4o-mini") # or any other tool-calling capable LLM
search_tool = TavilySearchResults(max_results=3)
tools = [search_tool]

# Create the prebuilt ReAct agent
agent_executor = create_react_agent(llm, tools=tools)

# Example usage
# Ensure OPENAI_API_KEY and TAVILY_API_KEY are set
if os.environ["OPENAI_API_KEY"] == "YOUR_OPENAI_API_KEY" or os.environ["TAVILY_API_KEY"] == "YOUR_TAVILY_API_KEY":
    print("Please set OPENAI_API_KEY and TAVILY_API_KEY environment variables or replace placeholders.")
else:
    response = agent_executor.invoke(
        {"messages": [("human", "What is the weather in London and what is 123 + 456?")]}
    )
    print(response["messages"][-1].content)