Agentic formative assessment for object oriented programming through multi source evidence aggregation

Abstract

Current AI-based programming education assessment systems do not involve process-oriented learning and a multi-faceted evidence of learning other than code correctness. This paper presents a three-partnering model of AI-Teacher, AI-Student, and Aggregator agents that combines multiple sources of evidence (syntax, semantics, process, and behavior) into a centralized learner state for formative OOP assessment. The framework evaluated on 2,550 student submissions has a larger grading accuracy (Cohen’s \(\kappa = 0.90\) ), adequate error coverage ( \(85.0\%\) ), high explanation quality (4.6/5), and substantial simulated learning gains ( \(26.1\%\) after six feedback cycles), surpassing traditional autograders and large-effect single-agent LLMs. Handling 1, 710 submissions per hour ( \(67\%\) points faster than LLM-only) the framework is a scalable, pedagogically valuable option in automated formative assessment in programming education.