A Code-Agency-Centered Evidence Framework for Programming Education in the Age of AIGC

Abstract

Generative artificial intelligence (AIGC) has made executable code easier to produce, but it has also weakened the evidential link between submitted code and students' actual programming competence. In programming education, the visibility of code products now contrasts with the relative invisibility of competence formation: students may submit code that runs without being able to explain its logic, judge its risks, revise its defects, or take responsibility for its consequences. This paper develops a Code-Agency-centered evidence framework for AIGC-supported programming education. Code Agency is deliberately bounded as educational ownership over code, expressed through four capacities: understanding code logic, judging code quality and risk, modifying code under constraints, and assuming responsibility for code behavior. The framework organizes learning through three progressive stages--independent implementation, AI-mediated critique and revision, and responsible engineering delivery--while treating AI access as a governance condition coupled with evidence requirements. The evidence cycle comprises execution-model construction, debugging and verification, AI output review and revision, and on-site defense with feedback adjustment. A LiteOS task-management case is used as a diagnostic setting because RTOS runtime states are partly invisible, scheduling behavior is temporally complex, and AI-generated code may appear correct while remaining poorly understood. The framework contributes a bounded and assessable account of code ownership for programming education in the age of AIGC.