How they actually differ
LangChain and Pydantic AI both let you build LLM agents but they take opposite stances on what an SDK should do, and the right pick depends mostly on how much abstraction you want sitting between your code and the model.
LangChain is a maximalist framework. It ships abstractions for almost every part of an LLM app: chains, agents, memory, retrievers, output parsers, tools, callbacks, an enormous integration library. That breadth is the value when you are prototyping fast and the cost when you need to debug something three layers deep. LangGraph (the same team's graph-runtime extension) is genuinely good for stateful agent workflows and is what we usually point teams to for serious production agent work.
Pydantic AI is a minimalist framework from the Pydantic team. Type-safe tool definitions, structured outputs as the default, agents as plain Python classes, and almost no hidden control flow. If you already use Pydantic for validation (most modern Python codebases do), the ergonomics feel native. Smaller integration surface, fewer batteries-included features, faster to read end-to-end.
Where the trade-off bites: LangChain is the right pick when the integration breadth saves you time (retrieval over five vector stores, ten provider SDKs, ready-made tools) and when LangGraph's state machine matches your control flow. Pydantic AI is the right pick when you want to write the control flow yourself, value strong typing, and prefer to keep the call stack legible.
Both work with Respan. We auto-instrument both via the Pydantic AI integration and the LangGraph integration. Spans land in the same trace tree so you can run side-by-side experiments without switching observability tools. See agent workflow patterns for which patterns each framework handles best.
