The virtuous loop

Observability and evaluation form a loop: traces show you what's happening and surface failures; you turn notable cases into eval examples; evals catch regressions when you change things; production signals reveal new gaps; those become new eval cases. An agent without this loop drifts; every change is a gamble. An agent with it improves steadily and safely. For probabilistic systems, this discipline isn't optional polish; it's how you engineer quality at all.