Are Your Agents Ready to Evolve? Assessing Evolution-Readiness in Agents Through Paired Safety and Utility Scoring

As AI agents gain the ability to adapt by accumulating memories, saving reusable skills, revising prompts, and delegating to subagents, they are generally expected to improve with experience. Enabling that ability presupposes evolution-readiness, the assumption that what an agent learns from one task will leave it no less safe and no less capable on the next. Existing agent-safety evaluations hold an agent's persistent state fixed and score standalone tasks, so they cannot test this assumption. Therefore, we introduce EvoAudit, a task suite purpose-built for agent evolution that uses directed source-target task pairs and matched cold and evolved conditions to isolate how a prior experience changes subsequent behavior. Every task is scored jointly on safety and utility, and the audit spans benign and dangerous tasks across three domains: DevOps, FinTech, and Workplace. Using a self-evolving agent framework, we run controlled experiments across five models and seven evolution mechanisms, covering memory, skills, prompt rewriting, subagent delegation, and their automatically injected meta variants. Evolution leaves behavior unchanged on most pairs, but where it does move behavior, misevolution is at least as frequent as improvement, and both tradeoff outcomes sacrifice more on the axis they give up than ordinary degradation or improvement gains. Across pair types, the improvement rate is stable; what varies is the form failure takes, with dangerous targets shifting joint degradation toward explicit safety-utility tradeoffs and enlarging the safety losses that accompany utility gains. Which outcome occurs is not attributable to the evolution mechanism or the backing model, so readiness cannot be secured by selecting either. It has to be established per trajectory and on both axes, and our results indicate that self-evolving agents remain some distance from the readiness that deploying them unsupervised would require.