I built a few agents with this last week. The orchestration layer is solid, but it still feels like we're just wrapping LLM calls in more abstractions.