LangChain Expert Skill

LangChain is an MIT-licensed Python framework for building LLM-powered applications. Since v1.0 (October 2025), it provides a layered architecture: high-level chain composition via LCEL (LangChain Expression Language), agent creation via create_agent (running on the LangGraph runtime underneath), and production observability via LangSmith. With 1000+ integrations and 100K+ GitHub stars, it is the most widely adopted LLM orchestration framework.

Key v1.0 change: All new LangChain agents run on the LangGraph runtime. AgentExecutor is in maintenance mode until December 2026. Use create_agent for new agents. Drop to LangGraph directly when you need full state-machine control.

⚠️ CRITICAL: Do NOT use AgentExecutor for new code. It is in maintenance mode until December 2026. Use create_agent(model, tools, prompt) instead — it generates a LangGraph state machine with streaming, persistence, and observability out of the box.

Core Principles

These principles govern every decision when building with LangChain. Read them before proceeding to the reference guides.

1. LCEL is the composition primitive. The pipe operator (|) chains Runnables. Every component — prompt, model, parser, retriever — implements the Runnable interface. Build everything in LCEL.
2. Agents run on LangGraph. Since v1.0, create_agent generates a LangGraph state machine underneath. You get streaming, persistence, and observability without writing graph code. Drop to LangGraph when you need branching, cycles, or human-in-the-loop.
3. RAG is a chain, not a framework. retriever | prompt | model | parser is the canonical RAG pattern. Document loaders, splitters, and vector stores are all interchangeable components.
4. LangSmith is production observability. Enable tracing at startup. 89% of production teams use observability — without it, debugging agent behavior is guesswork.
5. The ecosystem is the moat. 1000+ integrations mean model providers, vector stores, and tools are swappable with one line. Build against the interface, not the implementation.

Where to Start