LangGraph

What it is: LangGraph is LangChain's lower-level framework for building stateful, multi-step agents as explicit graphs. You define nodes (functions or LLM calls) and edges (transitions), and LangGraph handles persistence, branching, and human-in-the-loop checkpoints.

Why it matters for data work

Most real data agent work isn't a single prompt — it's "fetch, validate, transform, retry on failure, ask the user, write back". A graph framework lets you express that as code instead of nested if-statements buried in a system prompt.

Install & configure

pip install langgraph

You define a StateGraph with typed state, register nodes, wire edges (including conditional edges), and compile. Persistence backends include in-memory, SQLite, and Postgres for durable long-running runs.

Example usage

Common pattern: build an ETL agent where the graph fetches a CSV, runs a pydantic validator node, branches to a "fix or skip" subgraph on failures, then writes to a warehouse — with checkpoints between every step so a transient failure doesn't lose hours of work.

Author & links

Author: LangChain

Repo: github.com/langchain-ai/langgraph

License: MIT

Related skills

For a more typed, less-batteries-included approach see Pydantic-AI.

← Back to Agent Frameworks