LLM Frameworks: Raw SDK, LangChain, LangGraph and Agent SDKs
What each layer of the LLM tooling stack gives you, what it costs in complexity, and how to choose for a project.
AI ToolsIntermediate2 min readUpdated 11 Oct 2026
The layers
| Layer | Gives you | Cost |
|---|---|---|
| Provider SDK (direct API) | Full control, fewest surprises | You write retries, tool loops and state yourself |
| Component libraries (e.g. LangChain) | Integrations: loaders, vector stores, model wrappers | Abstractions to learn; version churn |
| Graph / workflow frameworks (e.g. LangGraph) | Explicit state machines, checkpoints, human-in-the-loop | More structure to design |
| Agent SDKs from model providers | Tool loops, handoffs, tracing out of the box | Tied to one ecosystem's conventions |
How to choose
- A single call or a simple RAG endpoint: the provider SDK is often clearest.
- Many integrations (loaders, stores) you would otherwise write: a component library.
- Multi-step workflows that must resume, branch and pause for approval: a graph framework.
- Tool-using agents where built-in tracing and handoffs fit your needs: an agent SDK.
Keep your core logic portable
Put prompts, tool functions and business rules in your own modules. Let the framework call them. Then switching frameworks, or removing one, is a contained change.
# your code: framework-independent
def lookup_order(order_id: str) -> dict: ...
# thin adapter for whichever framework you use
tools = [as_framework_tool(lookup_order)]Finished this note?