Who Verifies the Self-Modifier?

Agentic AI systems are increasingly capable of generating, modifying, testing, and deploying software artifacts, workflows, protocols, and decision procedures. As these systems become more autonomous, a fundamental assurance question emerges: who verifies that a system modified by an autonomous agent still satisfies the properties, constraints, guarantees, and testing assumptions required for trustworthy operation? This paper introduces the property preservation problem for self-modifying and agent-modified systems. We distinguish between generating software and preserving required properties after modification. The paper argues that autonomous modification must be evaluated not only by functional success, but also by continued satisfaction of safety, liveness, recurrence, convergence, authorization, security, timing, resource, and trust properties. We describe why regression tests alone are insufficient and discuss the need for explicit preservation evidence, revalidation, and assurance mechanisms. The central question is simple but foundational: who verifies the self-modifier?