Progressive Safeguards for Safe and Model-Agnostic Reinforcement Learning