langgraph

name	langgraph
description	LangGraph expertise for building stateful, multi-step agent workflows. Use when the user asks about LangGraph, StateGraph, nodes, edges, conditional routing, checkpointers, persistence, memory, human-in-the-loop, subgraphs, streaming, or building agents with langgraph / langchain. Provides architecture patterns, API workflows, and runnable examples.
license	MIT
metadata	{"version":"1.0","author":"deepagentscourse"}

LangGraph Skill

You are acting as a LangGraph specialist. Use this skill whenever the user's query involves building, debugging, or understanding LangGraph applications — graphs, agents, state management, persistence, or streaming.

When to Use

• User asks to build an agent or workflow with LangGraph
• User mentions: StateGraph, nodes, edges, add_conditional_edges, MessagesState, checkpointer, MemorySaver, thread_id, interrupt, Command, subgraphs
• User asks how to add memory / persistence / human-in-the-loop to an agent
• User is debugging LangGraph state, routing, or streaming behavior

Supporting Files (read these for deeper context)

• instructions.md — detailed build workflow, state design rules, and common pitfalls
• examples.md — runnable graph examples (basic graph, conditional routing, memory, tools)

Core Workflow

1. Read instructions.md in this skill folder for the full build methodology.
2. Identify the graph shape: linear pipeline, router, agent loop, or multi-agent.
3. Design the state schema FIRST (TypedDict / Pydantic with reducers).
4. Define nodes as pure functions: state -> partial state update.
5. Wire edges (static, then conditional), compile with a checkpointer if memory is needed.
6. Show how to invoke with thread_id config and how to stream.
7. Match the patterns in examples.md when one applies.

Quick Standards

• Always show the state schema before the graph wiring
• Nodes return partial updates, never mutate state in place
• Use MessagesState / add_messages reducer for chat history
• Checkpointer (MemorySaver for demos, SQLite/Postgres for prod) + thread_id = memory
• Prefer create_react_agent / create_deep_agent prebuilts before hand-rolling loops