Building Autonomous Multi-Agent Workflows with LangGraph: State, Nodes & Conditional Routing

Orchestrating Complex Enterprise Tasks via Cyclic Multi-Agent Architectures

While linear LLM pipelines (such as standard LangChain chains or sequential prompts) work well for basic question-answering, real-world business automation demands cyclical reasoning: an agent drafts code, another agent executes and unit-tests it, and if errors occur, the state loops back to the authoring agent with stack trace feedback. LangGraph models multi-agent workflows as stateful directed cyclic graphs, granting developers granular control over agent state transitions, human-in-the-loop approvals, and rollback persistence.

Core Concepts: State, Nodes, and Conditional Edges

• TypedDict State: A shared memory context updated by agent nodes across the execution lifecycle.
• Nodes: Individual worker functions or LLM agents (e.g., Researcher, Writer, Financial Reviewer) that consume the state and return delta updates.
• Edges: Conditional logic functions determining the next step: route to another node, retry the task, or terminate execution at END.

Complete Python Code Example: Financial Auditor Agent Loop

from typing import TypedDict, Annotated, Sequence
import operator
from langgraph.graph import StateGraph, END

class AgentState(TypedDict):
    draft_analysis: str
    audit_feedback: str
    is_approved: bool
    iterations: int

def analyst_node(state: AgentState):
    iter_count = state.get("iterations", 0) + 1
    return {
        "draft_analysis": f"Revised financial commentary version {iter_count}",
        "iterations": iter_count
    }

def auditor_node(state: AgentState):
    if state["iterations"] >= 2:
        return {"audit_feedback": "Passed all balance tests.", "is_approved": True}
    return {"audit_feedback": "Discrepancy in Q3 Net Profit calculation.", "is_approved": False}

def route_next(state: AgentState):
    if state["is_approved"] or state["iterations"] >= 3:
        return END
    return "analyst"

workflow = StateGraph(AgentState)
workflow.add_node("analyst", analyst_node)
workflow.add_node("auditor", auditor_node)

workflow.set_entry_point("analyst")
workflow.add_edge("analyst", "auditor")
workflow.add_conditional_edges("auditor", route_next)

app = workflow.compile()
final_state = app.invoke({"draft_analysis": "", "audit_feedback": "", "is_approved": False, "iterations": 0})
print("Workflow Final Result:", final_state)