We investigate how agent harness evolution can achieve recursive self‑improvement (RSI) without altering model weights. The central question is whether higher benchmark scores translate into reliable, efficient and transferable gains. Four methods are benchmarked on two harness families—general‑task and coding‑task—using repeated evaluations, controlled agent expansion, component interventions and admission replays. A consistent pattern emerges: all sixteen evolved agents acquire previously unseen tasks while losing tasks they had mastered. In one general‑task scenario reliability rises by 5.2 points but ten previously covered tasks disappear. Most reliability gains come with higher deployment cost. Targeted modifications usually outperform generic reminders, yet their benefit may flip when combined. The top‑scoring agent does not deliver the largest mean improvement in any pool, and stricter admission criteria can discard useful alternatives. Transferability is selective: coverage typically drops on the general‑task target across both harness types, but rises on the coding target when using an outer‑layer‑only harness. These observations motivate a preservation‑aware evaluation framework, tested decision rules, and explicit applicability conditions. As a constructive follow‑up, EvoAudit converts these diagnostics into a four‑step pipeline—select, analyze, mutate, gate—achieving the best performance. Review: The study highlights that evaluating agent evolution requires balancing task coverage, cost, and transferability; optimizing a single metric may mask critical regressions.