Source note
From Code-Centric to Intent-Centric Software Engineering: A Reflexive Thematic Analysis of Generative AI, Agentic Systems, and Engineering Accountability
Summary
This qualitative study argues that GenAI and coding agents are shifting software engineering toward intent specification, context curation, verification, and governance. The paper is relevant to agentic software engineering, but it reports an interpretive analysis rather than a new model or benchmark result.
Problem
- GenAI lowers the cost of producing plausible code, which can increase hidden technical debt, security risk, and unclear accountability when teams adopt it for speed alone.
- Existing studies cover code generation, AI pair programming, and software-engineering agents, but this paper targets how public technical discourse and peer-reviewed evidence frame the profession’s near-term change.
- The key issue is how engineers should control human-agent workflows that span requirements, repository context, tests, build systems, scanners, deployment, and telemetry.
Approach
- The paper uses reflexive thematic analysis as the main method, with interpretative phenomenological analysis as a secondary lens for role, identity, authorship, and responsibility.
- The corpus has 3 evidence layers: peer-reviewed software engineering and AI research, technical preprints and benchmark artifacts, and public discourse such as talks, interviews, essays, product posts, and X-originated posts.
- The analysis uses a corpus register, codebook, coding matrix, theme-to-source traceability table, DOI/reference audit, reproducibility protocol, and SHA-256 manifest.
- The workflow includes semantic coding, latent coding, idiographic profiles for named researchers and practitioners, cross-case theme construction, triangulation against peer-reviewed literature, and claim auditing.
- The paper treats public thought leadership as discourse data, not as proof of technical performance.
Results
- The paper reports no quantitative model-performance result, benchmark score, user-study effect size, or productivity measurement of its own.
- It structures the study around 3 research questions: the shift from code production to intent, verification, and governance; the main practitioner dichotomies; and maturity stages for future roles and practices.
- It claims 4 contributions: a bounded qualitative account, a reproducible public-discourse corpus protocol, a dichotomy-centered interpretation, and a maturity-stage pathway with a research agenda.
- It defines 7 initial code families: programming interface, agentic workflow, architecture and context, verification and evidence, professional identity, governance and risk, and capability skepticism.
- Its main claim is that engineers will spend less effort on first-draft code and more effort on specifying intent, curating repository context, supervising agents, validating outputs, preserving architecture knowledge, and enforcing governance.
- The strongest concrete process claim is a 10-step research workflow covering corpus definition, inclusion criteria, evidence capture, semantic and latent coding, idiographic profiles, cross-case themes, triangulation, and claim audit.