Langgraph

LangGraph Skill

When to activate

• Building a stateful agent that needs to pause, resume, or branch based on conditions
• Implementing human-in-the-loop workflows (agent pauses for approval before continuing)
• Orchestrating multi-step agents with complex branching logic
• Persisting agent state across failures or sessions (checkpointing)
• Building multi-agent systems where subgraphs collaborate

When NOT to use

• Simple single-turn AI calls — use the Claude API or Vercel AI SDK skill directly
• Linear pipelines with no branching — overkill, use sequential function calls
• When you need a hosted agent loop — use the Mastra skill instead
• Frontend streaming chat UIs — use the Vercel AI SDK skill

Instructions

Installation

bashCopy code
npm install @langchain/langgraph @langchain/anthropic @langchain/core
# Python
pip install langgraph langchain-anthropic

Core concepts

LangGraph models agents as graphs:

• Nodes — functions that run and update state
• Edges — connections between nodes (always run → next node, or conditional)
• State — a typed object that flows through the graph and accumulates updates
• Checkpointer — persists state between runs (enables resume, human-in-the-loop)

Basic StateGraph (TypeScript)

typescriptCopy code
import { StateGraph, Annotation, END, START } from '@langchain/langgraph'
import { ChatAnthropic } from '@langchain/anthropic'
import { tool } from '@langchain/core/tools'
import { z } from 'zod'

// 1. Define state shape
const AgentState = Annotation.Root({
  messages: Annotation<BaseMessage[]>({
    reducer: (existing, update) => [...existing, ...update],
    default: () => [],
  }),
  toolResults: Annotation<string[]>({
    reducer: (existing, update) => [...existing, ...update],
    default: () => [],
  }),
})

// 2. Define tools
const searchTool = tool(
  async ({ query }) => `Search results for: ${query}`,
  { name: 'search', description: 'Search the web', schema: z.object({ query: z.string() }) }
)

// 3. Define nodes
const model = new ChatAnthropic({ model: 'claude-opus-4-7' }).bindTools([searchTool])

async function callModel(state: typeof AgentState.State) {
  const response = await model.invoke(state.messages)
  return { messages: [response] }
}

async function callTools(state: typeof AgentState.State) {
  const lastMessage = state.messages[state.messages.length - 1]
  const results = []
  for (const toolCall of lastMessage.tool_calls ?? []) {
    const result = await searchTool.invoke(toolCall.args)
    results.push(new ToolMessage({ content: result, tool_call_id: toolCall.id }))
  }
  return { messages: results }
}

// 4. Conditional edge — should we keep looping?
function shouldContinue(state: typeof AgentState.State) {
  const lastMessage = state.messages[state.messages.length - 1]
  if (lastMessage.tool_calls?.length) return 'tools'
  return END
}

// 5. Build the graph
const graph = new StateGraph(AgentState)
  .addNode('agent', callModel)
  .addNode('tools', callTools)
  .addEdge(START, 'agent')
  .addConditionalEdges('agent', shouldContinue)
  .addEdge('tools', 'agent')
  .compile()

// 6. Run it
const result = await graph.invoke({
  messages: [new HumanMessage('Research the latest AI news and summarise it')]
})

Persistence and checkpointing (resume after failure)

typescriptCopy code
import { MemorySaver } from '@langchain/langgraph'
// Or for production: PostgresSaver, SqliteSaver

const checkpointer = new MemorySaver()

const graph = new StateGraph(AgentState)
  // ... nodes and edges ...
  .compile({ checkpointer })

// Run with a thread ID — state persists across calls
const config = { configurable: { thread_id: 'user-123-session-1' } }

// First call
await graph.invoke({ messages: [new HumanMessage('Start research')] }, config)

// Resume later (or after failure) — picks up from last checkpoint
await graph.invoke({ messages: [new HumanMessage('Continue')] }, config)

// Get current state
const state = await graph.getState(config)
console.log(state.values.messages)

Human-in-the-loop (interrupt before sensitive action)

typescriptCopy code
import { interrupt } from '@langchain/langgraph'

// Node that pauses for human approval
async function reviewBeforeSend(state: typeof AgentState.State) {
  const draft = state.messages[state.messages.length - 1]

  // Pause execution — returns control to the caller
  const approved = interrupt({
    message: 'Please review this draft before sending:',
    draft: draft.content,
  })

  if (!approved) {
    return { messages: [new HumanMessage('Draft rejected — please revise')] }
  }

  await sendEmail(draft.content)
  return { messages: [new AIMessage('Email sent successfully')] }
}

const graph = new StateGraph(AgentState)
  .addNode('draft', generateDraft)
  .addNode('review', reviewBeforeSend)  // will pause here
  .addNode('send', finaliseAndSend)
  .addEdge(START, 'draft')
  .addEdge('draft', 'review')
  .addEdge('review', 'send')
  .compile({ checkpointer, interruptBefore: ['review'] })

// Run until interrupt
const result = await graph.invoke(input, config)
// result.status === 'interrupted' — show the draft to user

// Resume after human approval
await graph.invoke(new Command({ resume: true }), config)
// Or reject:
await graph.invoke(new Command({ resume: false }), config)

Multi-agent subgraphs

typescriptCopy code
// Subgraph for research tasks
const researchGraph = new StateGraph(ResearchState)
  .addNode('search', searchNode)
  .addNode('extract', extractNode)
  .addNode('summarise', summariseNode)
  .addEdge(START, 'search')
  .addEdge('search', 'extract')
  .addEdge('extract', 'summarise')
  .addEdge('summarise', END)
  .compile()

// Main orchestrator graph
const orchestratorGraph = new StateGraph(OrchestratorState)
  .addNode('plan', planNode)
  .addNode('research', researchGraph)  // entire subgraph as a node
  .addNode('write', writeNode)
  .addNode('review', reviewNode)
  .addEdge(START, 'plan')
  .addEdge('plan', 'research')
  .addEdge('research', 'write')
  .addConditionalEdges('write', needsRevision, { yes: 'review', no: END })
  .addEdge('review', 'write')
  .compile({ checkpointer })

Python equivalent

pythonCopy code
from langgraph.graph import StateGraph, START, END
from langgraph.prebuilt import ToolNode
from langchain_anthropic import ChatAnthropic
from typing import TypedDict, Annotated
import operator

class AgentState(TypedDict):
    messages: Annotated[list, operator.add]

model = ChatAnthropic(model="claude-opus-4-7").bind_tools(tools)

def should_continue(state: AgentState) -> str:
    last = state["messages"][-1]
    return "tools" if last.tool_calls else END

graph = (
    StateGraph(AgentState)
    .add_node("agent", lambda s: {"messages": [model.invoke(s["messages"])]})
    .add_node("tools", ToolNode(tools))
    .add_edge(START, "agent")
    .add_conditional_edges("agent", should_continue)
    .add_edge("tools", "agent")
    .compile(checkpointer=MemorySaver())
)

result = graph.invoke(
    {"messages": [HumanMessage("Research AI news")]},
    config={"configurable": {"thread_id": "session-1"}}
)

Example

User: Build a content generation agent that: researches a topic, drafts an article, pauses for human review, then publishes — with state persisted so it can be resumed if interrupted.

Expected output:

• AgentState with messages, draftContent, approved, published fields
• Nodes: research, draft, review (with interrupt()), publish
• MemorySaver checkpointer, thread_id per content piece
• API endpoint: POST /generate (starts graph), POST /approve (resumes with Command({ resume: true }))
• State retrieval: GET /status/:threadId using graph.getState(config)