Evals are only as good as intent?

Most people build Evals in order to keep things a bit reliable when letting agents work autonomously. But even the best eval is still only as good as the formulated intent. The big issue I see is that most traditional software engineers have been struggling with getting the requirements right, unambiguous and closing the gap between "works as designed" and "works as intended." Now, a classic test case is binary: it fails if things aren't exactly as specified, so errors are fast and easy to detect. But even that isn't always reliable. And now we get AI agents, and the very people who didn't provide good requirements for decades - are now building a handful of evals for high-stakes automation, and "lgtm?" How do you practically assign responsibility for agentic behavior when: - the people who design it don't know how it needs to be tested? - the people who know how to design and test it aren't aware what is happening? - the people who are accountable have no clue what has been designed? - the people who have to bear the heat when it goes wrong never got a seat at the table? This is not a technology problem, and I don't think that "better automation" is the answer here.