Key points are not available for this paper at this time.
Abstract Large Language Model (LLM)-based software engineering excels at short tasks but fails on multi-day workflows due to memory loss, inconsistent reasoning, and lack of self-correction mechanisms. We propose VibeLayer Development (VLD), an architectural pattern that integrates existing technologies—persistent external storage, ordered event logging, LLM-based planning, and hybrid validation—into a unified six-layer framework. While each component is individually established, we argue that their specific integration achieves emergent properties unattainable by any subset: deterministic state reconstruction, long-term consistency, and closed-loop self-improvement. We present a formal model of VLD's state transitions, analyze the emergent properties arising from layer integration, and propose an evaluation framework with four metrics for assessing long-horizon agent stability. We hypothesize that VLD will significantly reduce task duplication and regression bugs compared to existing systems, and outline planned experiments to validate these claims.
Yusuke Harada (Thu,) studied this question.