Shift-Up: GenAI Software Engineering Framework

• Shift-Up is a framework for GenAI-native software development that reinterprets traditional software engineering artifacts as control mechanisms for agent-driven implementation.
• It integrates machine-readable requirements, architectural models, and decision records to stabilize agent behavior and reduce implementation drift.
• By reallocating human effort toward high-level design and validation, Shift-Up enhances traceability and mitigates the pitfalls of unstructured vibe coding.

Shift-Up is a framework for GenAI-native software development that reinterprets established software engineering practices as structural guardrails for agent-driven implementation. In this usage, the term denotes both a shift in abstraction level—from human involvement in code and low-level implementation toward engagement with system requirements, architecture, and system-level considerations—and an artifact-centered workflow in which machine-readable requirements, architectural models, and decision records are embedded directly into the development pipeline to stabilize agent behavior, reduce implementation drift, and shift human effort toward higher-level design and validation activities (Lipsanen et al., 22 Apr 2026, Stirbu et al., 29 Sep 2025).

1. Definition and problem setting

Shift-Up emerged in response to a reported change in software engineering practice in which Generative AI reshapes development from manual coding toward agent-driven implementation. Within that setting, the framework is positioned against “vibe coding,” which promises rapid prototyping but often suffers from architectural drift, limited traceability, and reduced maintainability. Shift-Up addresses those failure modes by treating traditional software engineering artifacts not as documentation produced after implementation, but as persistent control mechanisms that constrain and inform GenAI agents during implementation itself (Lipsanen et al., 22 Apr 2026).

A central premise is that the introduction of GenAI does not eliminate the need for structure; instead, it relocates where structure exerts force. The 2025 prestudy frames this as a “shift in abstraction level”: human teams concentrate on requirements engineering, architectural design, and system-level considerations, while detailed design, coding, and low-level verification are delegated to GenAI agents. A plausible implication is that Shift-Up is not merely a prompting technique, but a reallocation of cognitive and organizational effort across the software development lifecycle (Stirbu et al., 29 Sep 2025).

2. Research lineage and methodological basis

The framework is developed under Design Science Research (DSR). In the 2026 study, DSR is used to ground the artifact’s development in a rigorous, iterative process focused on practical problems such as agentic drift and lack of traceability, and to support systematic evaluation and refinement of the framework itself. The same methodological orientation appears in the 2025 prestudy, which also adopts DSR while testing the broader idea of GenAI-native development with human supervision concentrated on high-value activities (Lipsanen et al., 22 Apr 2026, Stirbu et al., 29 Sep 2025).

The two papers occupy different positions in the framework’s evolution. “Towards Shift-Up: A Framework and a Prestudy on High-Value Activities in GenAI Native Software Development” (Stirbu et al., 29 Sep 2025) proposes the general paradigm, emphasizes the V-Model as a basis for reallocating responsibilities, and reports a prestudy with five undergraduate interns developing a “travel history” web application using Claude, GPT, DeepSeek, and Gemini over approximately 10 person-months. “Shift-Up: A Framework for Software Engineering Guardrails in AI-native Software Development -- Initial Findings” (Lipsanen et al., 22 Apr 2026) operationalizes that agenda more narrowly through explicit guardrails—BDD, C4, and ADRs—and compares Shift-Up to unstructured vibe coding and structured prompt engineering in a web application case study for an online snack-bar. This suggests a progression from a general human–GenAI division-of-labor model toward a concrete artifact-governed implementation framework.

3. Guardrail artifacts

Shift-Up’s core mechanisms are executable requirements, architectural modeling, and architecture decision records. In the framework, Behavior-Driven Development is used to transform user stories into executable acceptance tests; in the reported study, this was done with the Robot Framework in a “given-when-then” format. These artifacts are machine-readable guardrails: the agent must generate code that passes the acceptance tests, so the requirements function simultaneously as specification and automated validation (Lipsanen et al., 22 Apr 2026).

Architectural structure is supplied through the C4 Model, with architecture described at multiple levels of abstraction—context, container, component, and code. Architecture Decision Records document architectural tradeoffs and rationale as structured records. Together, C4 and ADRs provide explicit architectural constraints and contextual rationale, promoting architectural consistency while constraining the generative agent’s solution space. The 2025 prestudy generalizes the same principle under “Use of Structured Knowledge,” referring to requirements, design documents such as the C4 Model, ADRs, and knowledge graphs as foundational context that reduces hallucination and misalignment (Stirbu et al., 29 Sep 2025).

The framework therefore differs from prompt-only approaches in a specific way. The 2026 study states that, in contrast to probabilistic prompt optimizations, Shift-Up employs deterministic, prescriptive artifacts to constrain and shape the agent’s output. All artifacts are ingested by the agent as contextual input, so the agent reasons over structured documents rather than relying on unstructured prompt history. For an arXiv-literate audience, the important point is that Shift-Up treats software engineering artifacts as executable and machine-consumable state, not as passive reference material (Lipsanen et al., 22 Apr 2026).

4. Workflow and lifecycle reallocation

The reported Shift-Up workflow begins with stakeholder needs elicited through structured interviews and synthesized into a Software Requirements Specification. The SRS is decomposed into user stories and then into BDD-format test cases. C4 models and ADRs are generated as architectural scaffolding, and an implementation roadmap divides development into phases associated with specific acceptance tests and architectural tasks. For each phase, the agent receives the requirements, associated BDD tests, and relevant C4 and ADR artifacts; generates implementation artifacts; and is allowed to proceed only when all tests for that phase pass (Lipsanen et al., 22 Apr 2026).

This workflow aligns with the layered SDLC integration described in the prestudy. There, the top levels of the V-Model—requirements and architecture—are human-driven; the middle levels—detailed design, implementation, and unit/integration testing—are GenAI-driven; and the bottom levels—acceptance testing and monitoring—remain under human oversight with GenAI support. The same paper presents a shared responsibility model in which humans lead high-level, context-rich activities and GenAI tools automate detail-intensive and routine work. In the activity ownership table, coding is assigned to GenAI tools for code generation, while requirements and acceptance activities remain anchored in human definition, quality control, and monitoring (Stirbu et al., 29 Sep 2025).

A plausible implication is that Shift-Up should be understood less as a single pipeline than as a control architecture over the SDLC. Its distinctive claim is not that GenAI performs implementation, but that implementation is gated by explicit behavioral and architectural artifacts and by continuous validation.

5. Comparative evaluation and reported findings

The 2026 case study compares three development paradigms in a web application for an online snack-bar. Unstructured vibe coding uses freeform prompting without formal requirements or architecture; it is described as the fastest for prototype, but tightly coupled to tool-specific guardrails and offering little support for maintainability. Structured vibe coding introduces lightweight planning and context-tracking via prompts, but lacks persistent, machine-readable artifacts. Shift-Up introduces full integration of BDD requirements, C4 and ADR artifacts, and test-driven acceptance with GenAI agents responsible for implementation (Lipsanen et al., 22 Apr 2026).

The preliminary findings reported for Shift-Up are specific. Embedding machine-readable requirements and architectural artifacts is said to stabilize agent behavior and reduce implementation drift. The generative agent is described as enabled “to operate with increased independence during implementation tasks.” Most prompts in Shift-Up concerned progressing tasks or running automated tests rather than manual error correction. Human effort was redistributed: in Shift-Up, the majority of prompts related to process orchestration (62%), test execution (16%), and progression, whereas in structured vibe coding over half (52%) of prompts were spent manually diagnosing or fixing issues. The framework was slower to start because of the investment in capturing and formalizing requirements and models, but the reported effect was a shift from low-level debugging toward higher-level orchestration, validation, and strategic intervention (Lipsanen et al., 22 Apr 2026).

The 2025 prestudy reports a related but broader set of observations. Human oversight remained indispensable even in a GenAI-native process: teams had to plan and communicate project scope and design context, treat prompt engineering as an iterative core skill, work in small and testable increments, scrutinize outputs, and manage versions because GenAI could recommend outdated tools or libraries. Those findings are consistent with Shift-Up’s emphasis on structured context and continuous validation, although the prestudy did not yet present the same explicit artifact-governed pipeline (Stirbu et al., 29 Sep 2025).

6. Significance, limitations, and open research directions

Shift-Up’s main significance lies in its claim that traditional software engineering artifacts can serve as effective control mechanisms in AI-assisted development. In this view, BDD, C4, and ADRs are not legacy process overhead; they are machine-readable, persistent, and traceable guardrails that shape agent reasoning and provide enforceable boundaries. The framework also aligns with a shared responsibility model that shifts humans upward on the V-Model toward requirements, design, and validation, while GenAI assumes implementation-centric work (Lipsanen et al., 22 Apr 2026, Stirbu et al., 29 Sep 2025).

The empirical basis, however, is preliminary. The 2026 study notes that the reduction in agent drift was only partially observed and that the proof was limited by the project being relatively standard—a web snack-bar application. It also states that implementation of Shift-Up for the whole project was not yet concluded as of reporting. The 2025 prestudy similarly characterizes current adoption as ad hoc and calls for more systematic and mature methodologies for industry-grade reliability (Lipsanen et al., 22 Apr 2026, Stirbu et al., 29 Sep 2025).

The identified future directions are correspondingly methodological. They include systematization and maturity of GenAI integration, better understanding of the limitations and best practices of prompt engineering, study of human–GenAI interaction and developer flow state, controlled benchmarking of GenAI-in-the-loop against human-only workflows, and integration into existing workflows through structured context and continuous validation. Taken together, these directions indicate that Shift-Up currently functions as an early research program in AI-native software engineering: its core proposition is that formalized software engineering artifacts can be repurposed as operational guardrails for agentic development, but its generality, scalability, and external validity remain open empirical questions (Stirbu et al., 29 Sep 2025).