How to Build an AI Agent with LangGraph and Claude

LangGraph is quickly becoming the industry standard orchestration framework for multi-agent workflows. In this tutorial, we will build a robust AI researcher using Anthropic's Claude 3.5 Sonnet.

The Paradigm: State Machines

LangGraph represents agents as state machines (graphs). You define:

• State: The schema representing the state of your application.
• Nodes: Python functions representing agents or tools.
• Edges: The conditional logic connecting the nodes.

1. Defining the State

We define a dictionary-like state containing the conversation history.

from typing import TypedDict, Annotated
from langgraph.graph.message import add_messages

class AgentState(TypedDict):
    messages: Annotated[list, add_messages]

2. Defining Tools and the LLM

We bind a mathematical calculator tool to Claude.

from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_anthropic import ChatAnthropic

# An internet search tool
search_tool = TavilySearchResults(max_results=2)
tools = [search_tool]

# Initialize Claude and bind the tools to it
llm = ChatAnthropic(model="claude-3-5-sonnet-20240620")
llm_with_tools = llm.bind_tools(tools)

3. Creating the Graph Nodes

The two primary nodes: the LLM making choices, and the environment executing tools.

from langgraph.prebuilt import ToolNode

# Node 1: Executing the LLM
def invoke_agent(state: AgentState):
    response = llm_with_tools.invoke(state["messages"])
    return {"messages": [response]}

# Node 2: The pre-built tool executor
tool_node = ToolNode(tools)

4. Wiring the Graph

We define the control flow. If Claude calls a tool, go to the tool node. Otherwise, end.

from langgraph.graph import StateGraph, END
from langgraph.prebuilt import tools_condition

graph_builder = StateGraph(AgentState)

# Add nodes
graph_builder.add_node("agent", invoke_agent)
graph_builder.add_node("tools", tool_node)

# Add edges
graph_builder.set_entry_point("agent")

# If the agent calls a tool -> go to tools. Otherwise -> END.
graph_builder.add_conditional_edges("agent", tools_condition)
graph_builder.add_edge("tools", "agent")

# Compile the final graph
agent_executor = graph_builder.compile()

5. Execution

response = agent_executor.invoke({
    "messages": [("user", "Search the web for the latest developments in LK-99 superconductors and summarize them.")]
})

print(response["messages"][-1].content)

Claude will now iteratively plan searches, read the results from Tavily, and synthesize a final report.