My all-things-considered recommendation, both from talking to people and living my life, is that if you’re going to be in a codebase for longer than about 3 months you should roll your own agentdev abstractions. Frameworks are amazing for quick projects, but start to get on your nerves if you’re doing something seriously custom. Unless you’re like the platonic ideal of the inspect-ai user, though I don’t think formal methods evals really are. pydantic-ai has an even less academic primary usecase, so is probably even less appropriate.