I've built the same toy project (personalized bedtime story generator) with lang-chain and lang-graph + pydanticAI. I preferred the latter. With agents right now the priority is to try to wrestle predictability out of the models and working with pydantic backed dependencies and state is great