What it is

LangChain is an open-source framework for combining models with tools, retrieval, structured output, and agents. Use it when those integrations remove more complexity than they add; otherwise, prefer a direct model call.

Core pieces

• Models and messages normalize generation, tool calls, and conversation turns.
• Tools pair a callable operation with a name, description, and input schema.
• Retrievers return relevant documents for a query.
• Agents let a model choose and call tools in a bounded loop.
• Middleware adds guardrails, state, retries, or observability.

Current LangChain agents use LangGraph as their underlying graph runtime. Reach for LangGraph directly when the workflow needs explicit state, branching, durability, or human checkpoints.

Production habits

• Pin versions and keep tool schemas small.
• Enforce authentication and authorization inside tools.
• Bound steps, retries, time, and spend.
• Trace and evaluate complete outcomes; keep fixed steps deterministic.

Further reading: LangChain overview and retrieval.