The bug fix paradox: why AI agents keep breaking working code

AI coding agents often over-engineer bug fixes by modifying unrelated code, adding unnecessary defensive checks, and drifting from the original intent. Kiro addresses this with a methodology called property-aware code evolution, which formalizes a bug condition (C) defining exactly when a bug triggers, a postcondition (P) defining what correct behavior looks like, and two testable properties: a fix property (C ⟹ P) and a preservation property (not C ⟹ unchanged behavior). The workflow generates a bugfix document, derives a falsifiable root cause hypothesis, writes property-based tests using Hypothesis before any fix is applied, and only then writes the minimal patch. The approach is demonstrated with a BST delete crash and a RocketMQ memory leak, both resolved with one-line fixes validated by comprehensive property-based test suites.