Solid explanation. For multi-model chains like render pipelines, where do you add critic agents—after each tool call or end-of-loop?