Libraries that structure prompts, tools, retrieval, and multi-step workflows around LLM APIs—accelerating agents and RAG when the fit is right, costly when over-applied.
LangChain popularized composable chains and tool wiring; LangGraph adds stateful graphs, checkpoints, and parallel branches for agents; LlamaIndex centers document ingestion, indexing, and RAG-heavy retrieval flows.
Choosing vs custom code
Frameworks help when you need retries, streaming, routing, and observability hooks without rebuilding them. They hurt when the task is a thin wrapper around one API call—then prefer plain code and tests. Pair framework traces with LLM observability (for example LangSmith or Langfuse) when teams debug production flows.
See also
Model Context Protocol for portable tool servers, and prompt chaining as a pattern that can live inside or outside a framework.