LangGraph

LangGraph is LangChain's state machine layer — it uses LangChain's runnable interface, tools, and model connectors as its graph primitives.

→ Stateful, cyclical agent graphs built on LangChain's full ecosystem — every LangChain tool is a potential graph node.

LangGraph sends traces to LangSmith automatically when LANGCHAIN_TRACING_V2 is set — every node execution becomes a separate trace span.

→ Step-by-step graph execution visibility: see which nodes ran, in what order, with what inputs, outputs, and token cost.

LangGraph nodes call LiteLLM-proxied endpoints, routing each node's LLM calls to any provider without changing graph code.

→ Model-agnostic LangGraph agents — route different nodes to Claude, GPT-4o, or local models via one LiteLLM config.

LangGraph integrates with Langfuse via its callback system or OpenTelemetry, capturing every node execution as a nested trace span.

→ Full execution traces of complex agent graphs — cost per node, latency per step, and LLM call details in one view.

LangGraph routes code-execution tool calls to E2B sandboxes, giving code-generation nodes a safe isolated environment to run output.

→ Agentic code generation with sandboxed execution built into the graph — iterate on code safely without leaving the agent loop.

LangGraph retrieval nodes query pgvector via psycopg2 or SQLAlchemy, keeping vector search within the existing Postgres database.

→ Enterprise RAG with no separate vector DB — LangGraph agents retrieve from the same Postgres instance the rest of the app uses.