I think those wrappers could create some potentially complex workflow around LLM API, with various trees of decisions, integrations, eval, rankers, ratets, etc, and this is their added value.

Wrappers are a bit pejorative and reductive - everything is a wrapper around something else.