Who Verifies the Self-Modifier? Property Preservation in Agentic AI Systems
Deepinder Sidhu · 2026
Recent advances in agentic AI systems and AIassisted software development raise a fundamental question: if an autonomous system modifies its own software, algorithms, protocols, tests, or decision processes, how can it be established that the resulting system continues to satisfy its required properties and constraints? While much attention has been devoted to capability improvement and code generation, comparatively little attention has been devoted to preservation of correctness properties following autonomous modification. This note introduces the property preservation problem and argues that preservation of system properties and test adequacy may represent fundamental challenges for self-modifying agentic systems.