🦜 LangChain & LangGraph

Mental model

LangChain = integrations + building blocks (models, prompts, retrievers, tools). LangGraph = a low-level framework for stateful, controllable agent workflows as graphs, with persistence, streaming, and human-in-the-loop. In 2026, use LangGraph for anything agentic. Python is the most mature; LangGraph.js exists; in Java, see LangChain4j / Spring AI.

LangChain

• Chat models, messages, prompt templates, output parsers / structured output
• Tools (@tool), tool-calling models
• Retrievers, document loaders, text splitters, vector store integrations
• Runnables / LCEL composition (awareness)

LangGraph ⭐

• StateGraph: state schema, nodes, edges, conditional edges, reducers
• Prebuilt ReAct agent vs a custom graph
• Persistence / checkpointers (Postgres checkpointer), threads, memory (short-term vs long-term store)
• Human-in-the-loop: interrupt, approve/edit tool calls, time travel
• Streaming (tokens, updates, events)
• Subgraphs, multi-agent patterns (supervisor, swarm/handoffs)
• Deployment options (LangGraph Platform / self-host with FastAPI), LangSmith tracing
• Connecting to MCP servers (MCP adapters)

🧪 Labs (🟢 warm-up → 🟡 core → 🔴 hard → ⚫ boss)

• 🟢 LangChain Academy Intro to LangGraph modules 1–2
• 🟡 A state graph with conditional edges + a Postgres checkpointer
• 🔴 interrupt for approvals + time travel + streaming to orbit-stream
• ⚫ A supervisor multi-agent (UC11) invoked as an Orbit agent step

🧠 Cognitive tasks

• “What did the framework hide?”: compare it with your Go ReAct agent

🛰️ Orbit integration

• The orbit-agents service

Go deeper

🧩 AI & Agent Patterns · 🤖 Agents & Workflows

Resources

• langchain-ai.github.io/langgraph · python.langchain.com · LangChain Academy · LangChain YouTube