LangGraph

Learning Objectives

• Understand what LangGraph is and how it differs from LangChain's higher-level APIs
• Identify when to use LangGraph for production agent workflows
• Compare LangGraph vs. CrewAI and AG2 for multi-agent orchestration

What Is LangGraph?

LangGraph is a low-level orchestration framework for building, managing, and deploying long-running, stateful AI agents. Created by LangChain AI, LangGraph models agent logic as directed graphs — where nodes represent actions (LLM calls, tool use, human input) and edges define the flow between them, including conditional branching and cycles.

Both LangChain and LangGraph reached their v1.0 milestones in early 2026, with production adoption at companies like Uber, LinkedIn, and Klarna. While LangChain provides high-level abstractions for common patterns, LangGraph gives developers fine-grained control over agent behavior — making it the recommended framework for production-grade agent systems.

LangChain agents are now built on LangGraph underneath, so you can start with LangChain's simple APIs and drop down to LangGraph when you need more control.

Core Concepts

State Graphs

LangGraph's fundamental abstraction is the StateGraph — a graph where each node reads and writes to a shared state object:

from langgraph.graph import StateGraph, END
from typing import TypedDict

class AgentState(TypedDict):
    messages: list
    next_step: str

graph = StateGraph(AgentState)

# Add nodes (functions that transform state)
graph.add_node("research", research_function)
graph.add_node("write", write_function)
graph.add_node("review", review_function)

# Add edges (flow between nodes)
graph.add_edge("research", "write")
graph.add_conditional_edges("review", should_revise, {"revise": "write", "approve": END})

app = graph.compile()
result = app.invoke({"messages": ["Write a report on AI agents"]})

Durable Execution

LangGraph agents persist through failures. If a node fails mid-execution, the agent can resume from the last successful checkpoint rather than restarting from scratch. This is critical for long-running workflows where a restart would waste significant compute and time.

Human-in-the-Loop

LangGraph provides first-class support for human oversight:

• Interrupt before: Pause execution before a specific node and wait for human approval
• Interrupt after: Let a node execute, then pause for human review before continuing
• Edit state: Humans can modify the agent's state at any checkpoint before resuming

Memory Systems

LangGraph supports both short-term working memory (state that persists within a single agent run) and long-term persistent memory (state that survives across sessions). This enables agents that learn from past interactions and maintain context over extended workflows.

Deep Agents

Deep Agents is LangGraph's newer capability for building agents that can plan multi-step strategies, spawn subagents for parallel subtasks, and use file systems for complex work. Deep Agents combine planning, delegation, and execution in a single orchestration pattern.

Comparison: Agent Frameworks

Pricing

LangGraph is free and open source (MIT license). Related costs:

• LLM API calls: charged by the providers you use
• LangSmith (observability): free tier available; paid plans from $39/month
• LangGraph Platform (managed deployment): enterprise pricing

Strengths

• Full control: Graph-based design lets you define exactly how agents make decisions, loop, branch, and recover from errors
• Durable execution: Agents persist through failures — critical for production reliability
• Human oversight: First-class interrupt, review, and state editing at any point in the workflow
• LangChain ecosystem: Access to 800+ integrations, document loaders, and tool libraries
• Observability: Deep integration with LangSmith for tracing every step of agent execution
• Memory: Both short-term and long-term memory for agents that maintain context across sessions

Limitations & Considerations

• Steep learning curve: Graph-based thinking requires more upfront design than role-based frameworks like CrewAI
• More code: Defining graphs requires more boilerplate than CrewAI's declarative role/task/crew syntax
• LangChain coupling: While usable standalone, LangGraph works best within the LangChain ecosystem
• Newer at scale: v1.0 is recent — enterprise deployment patterns are still emerging

Best Use Cases

When to choose alternatives:

• Simple role-based agent teams → CrewAI (faster to set up, more readable)
• Conversational multi-agent workflows → AG2
• No-code workflow automation → n8n or Zapier
• Single-agent with tool use → LangChain ReAct agent (higher-level, simpler)

Key Takeaways

• LangGraph models agent logic as stateful directed graphs — giving developers full control over agent behavior with durable execution and human-in-the-loop checkpoints
• v1.0 reached early 2026; production use at Uber, LinkedIn, Klarna; LangChain agents are built on LangGraph underneath
• Deep Agents enable planning, subagent delegation, and file system use for complex multi-step tasks
• More powerful than CrewAI for complex workflows; steeper learning curve — choose based on your use case complexity
• Free and open source (MIT); costs are the underlying LLM API providers