ephizen / wiki

LangGraph

A state-machine library from the LangChain team for building controllable, stateful LLM agents as explicit graphs of nodes and edges.

Category
LLM & Agent Frameworks
Difficulty
Intermediate
When to use
You're building an agent with branching logic, retries, checkpoints, or human-in-the-loop steps and want explicit control over the flow.
When not to use
A single LLM call with one tool is enough — LangGraph is overkill for one-shot prompts.
Alternatives
CrewAI AutoGen DSPy Custom state machine
https://www.langchain.com/langgraph →

At a glance

FieldValue
CategoryAgent / workflow orchestration
DifficultyIntermediate
When to useStateful agents, branching, human-in-the-loop
When not to useSimple one-shot LLM calls
AlternativesCrewAI, AutoGen, custom state machines

What it is

LangGraph models an agent as a directed graph. Nodes are functions (or LLM calls) that read and write a typed shared state; edges decide which node runs next based on the state. It supports checkpoints (pause and resume across machines), streaming, and human approval steps. It’s much closer to a workflow engine than the implicit loops of the original LangChain agents.

When we reach for it at Ephizen

  • Multi-step research agents with tool use and retry logic.
  • Approval flows where a human sits in the loop before a sensitive action runs.
  • Long-running workflows that need to survive process restarts via checkpoints.
  • Agents that branch on classification (“is this a support ticket or a bug report?”).

Getting started

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

class State(TypedDict):
    question: str
    answer: str

def answer_node(state: State) -> State:
    return {"answer": f"You asked: {state['question']}"}

graph = StateGraph(State)
graph.add_node("answer", answer_node)
graph.set_entry_point("answer")
graph.add_edge("answer", END)
app = graph.compile()
app.invoke({"question": "hello"})

Gotchas

  • The state schema matters — keep it small and serializable or checkpoints get expensive.
  • Debugging a large graph is painful without LangSmith tracing; set it up early.
  • Infinite loops are possible if edge conditions aren’t exhaustive. Always have a fallback edge to END.

Related tools