Agentic Interpretability: A Strategy Against Gradual Disempowerment — LessWrong