The Rise of AI-Native Software Engineering: Implications for Practice, Education, and the Future WorkforceAI原生软件工程的兴起：对实践、教育及未来劳动力的影响

Software engineering has evolved through successive waves of methodological and technological change, from structured programming and object-oriented design to agile delivery, DevOps, cloud-native development, and platform engineering. Each of these shifts has altered the instruments of practice, the organization of work, and the competencies expected of practitioners, yet across all of them a stable core endured: human developers authored, reasoned about, and maintained source code, while machines compiled, executed, and tested what people wrote. The emergence of generative artificial intelligence (GenAI), large language models (LLMs), and increasingly agentic systems marks a qualitatively different transition. These technologies do not merely augment individual tasks; they begin to reshape the distribution of cognitive labor across the software development lifecycle, including specification, coding, testing, debugging, documentation, maintenance, and release management. Code-capable assistants such as GitHub Copilot moved from research prototype to mainstream professional practice within roughly two years, and autonomous agents now attempt to resolve real repository issues end to end—reframing the operative question from “can a model complete a line of code?” to “can a system deliver a working, reviewable change?” Recent studies and roadmapping efforts increasingly characterize this transition as one of the most consequential changes in software engineering in decades [1, 34].软件工程经历了方法论和技术变革的连续浪潮，从结构化编程和面向对象设计，到敏捷交付、DevOps、云原生开发和平台工程。每一次转变都改变了实践工具、工作组织方式以及对从业者的能力要求，但贯穿其中的核心始终保持稳定：人类开发者编写、推理并维护源代码，而机器负责编译、执行和测试人类编写的内容。生成式人工智能（GenAI）、大语言模型（LLMs）以及日益强大的智能体系统的出现，标志着一种质的转变。这些技术不仅增强了单个任务，还开始重塑软件开发生命周期中认知劳动的分配，包括规范说明、编码、测试、调试、文档编写、维护和发布管理。GitHub Copilot等具备代码能力的助手在约两年内从研究原型转变为主流专业实践，而自主智能体现在正尝试端到端地解决真实的仓库问题——将核心问题从“模型能补全一行代码吗？”重新定义为“系统能交付一个可运行、可审查的变更吗？”近期的研究和路线图规划工作越来越多地将这一转变描述为几十年来软件工程领域最重大的变革之一 [1, 34]。

This transformation is especially significant because it affects software engineering simultaneously as a practice domain, an educational domain, and a labor-market domain. At the level of practice, AI-enabled tools are increasingly capable of generating code, proposing fixes, supporting test creation, and assisting with repository-level problem solving, as illustrated by recent benchmarks and agentic frameworks such as SWE-bench and SWE-agent [20, 47]. At the level of education, the same technologies are altering what students are expected to learn, how they are assessed, and which learning outcomes remain meaningful in an AI-mediated environment. At the level of the workforce, they are changing expectations for entry-level competence, professional identity, and the relative value of coding fluency versus judgment, verification, orchestration, and human–AI collaboration. These three dimensions are interdependent: if entry-level coding becomes increasingly automatable, then what universities assess, what employers hire for, and what the professional identity of a “software engineer” means must all change together. Any account that treats them separately therefore risks missing the systemic nature of the change.这种转型尤为重要，因为它同时影响了软件工程作为实践领域、教育领域和劳动力市场领域。在实践层面，AI赋能的工具正变得越来越强大，能够生成代码、提出修复建议、支持测试创建并辅助仓库级别的难题解决，正如SWE-bench和SWE-agent等近期基准测试和智能体框架所展示的那样 [20, 47]。在教育层面，同样的技术正在改变学生应学习的内容、评估方式，以及在AI中介环境下哪些学习成果仍具有意义。在劳动力层面，它们正在改变对入门级能力、职业身份以及编码流利度与判断力、验证、编排及人机协作相对价值的预期。这三个维度是相互依存的：如果入门级编码变得越来越可自动化，那么大学评估的内容、雇主招聘的标准以及“软件工程师”的职业身份定义就必须同步改变。因此，任何将它们分开对待的论述都有可能忽略这一变革的系统性本质。

Despite the rapid diffusion of these tools, the empirical evidence base remains fragmented and, in some respects, contradictory. While some studies report substantial productivity gains, improved developer experience, and accelerated task completion, others find limited benefits, increased overhead, or quality and security trade-offs that depend strongly on task type, developer expertise, and organizational context. The literature also reveals a tension between automation and accountability: the more capable AI systems become at producing code or executing routines, the more important human oversight becomes in validating correctness, robustness, and maintainability. Compounding this, the field is young, fast-moving, and dominated by preprints, so conclusions can shift between the time evidence is generated and the time it is formally reviewed. These inconsistencies make it difficult for educators, curriculum designers, and institutional leaders to draw stable conclusions about what should change in software engineering education and professional preparation.尽管这些工具普及迅速，但实证证据基础仍然支离破碎，且在某些方面存在矛盾。虽然一些研究报告称生产力大幅提升、开发者体验改善且任务完成速度加快，但另一些研究发现收益有限、额外开销增加，或者存在取决于任务类型、开发者经验和组织背景的质量与安全权衡。文献还揭示了自动化与问责制之间的张力：AI系统在生成代码或执行例行程序方面越强大，人类在验证正确性、稳健性和可维护性方面的监督就越重要。此外，该领域尚处于起步阶段且发展迅速，预印本占据主导，因此结论可能会在证据生成到正式评审期间发生变化。这些不一致性使得教育工作者、课程设计者和机构领导者难以就软件工程教育和职业准备中应做出哪些改变得出稳定的结论。

The implications for higher education are particularly urgent. If the routine production of syntactically correct code becomes increasingly automatable, then software engineering curricula can no longer rely on code generation as the primary proxy for competence. Instead, universities must reconsider the balance among problem framing, computational thinking, specification writing, verification, debugging, ethical reasoning, and collaboration with AI systems. This shift also raises important questions about assessment integrity, academic honesty, and the design of learning experiences that remain robust in an AI-rich environment. At the same time, faculty members require new forms of support to adapt teaching materials, evaluation strategies, and pedagogical models to a rapidly changing technical landscape.这对高等教育的影响尤为紧迫。如果语法正确的代码的常规生产变得越来越可自动化，那么软件工程课程就不能再依赖代码生成作为能力的主要代理指标。相反，大学必须重新审视问题框架、计算思维、规范说明编写、验证、调试、伦理推理以及与AI系统协作之间的平衡。这种转变还引发了关于评估完整性、学术诚信以及在AI丰富环境下如何设计仍具稳健性的学习体验等重要问题。同时，教职员工需要新型支持，以适应快速变化的技术环境，调整教学材料、评估策略和教学模型。

Against this backdrop, the present study systematically synthesizes the evidence from 2016 to 2026 to inform the design of AI-native software engineering education and practice. We ask how GenAI, LLMs, and agentic systems are changing software engineering work; what the empirical evidence indicates regarding productivity, quality, security, and collaboration; how education is responding in terms of pedagogy, tooling, and assessment; how professional roles and skill demands are evolving; and what competency model, curriculum roadmap, and research agenda follow from the accumulated evidence. To address these questions, we analyze a verified corpus of influential peer-reviewed studies and organize the findings into a coherent framework spanning practice, education, and workforce transformation. Concretely, the review is guided by five research questions:在此背景下，本研究系统地综合了2016年至2026年的证据，为AI原生软件工程教育和实践的设计提供参考。我们探讨了GenAI、LLMs和智能体系统如何改变软件工程工作；实证证据在生产力、质量、安全性和协作方面有何指示；教育在教学法、工具和评估方面如何应对；职业角色和技能需求如何演变；以及基于积累的证据，应得出什么样的能力模型、课程路线图和研究议程。为解决这些问题，我们分析了一个经过验证的有影响力的同行评审研究语料库，并将研究结果组织成一个涵盖实践、教育和劳动力转型的连贯框架。具体而言，本综述由五个研究问题引导：

• •

  RQ1. How have GenAI, LLMs, and agentic systems changed software engineering practice and which tasks are most affected?RQ1. GenAI、LLMs和智能体系统如何改变了软件工程实践，哪些任务受到的影响最大？

• •

  RQ2. What is the empirical evidence on developer productivity, code quality/security, and human–AI collaboration, and how consistent is it?RQ2. 关于开发者生产力、代码质量/安全性和人机协作的实证证据是什么，其一致性如何？

• •

  RQ3. How is computing/SE education responding—pedagogy, tools, and assessment—and what does the evidence show about learning effects?RQ3. 计算/SE教育如何应对（教学法、工具和评估），证据显示学习效果如何？

• •

  RQ4. How are professional roles, skills demand, and the engineer’s identity evolving?RQ4. 职业角色、技能需求和工程师的身份如何演变？

• •

  RQ5. What competency model, curriculum roadmap, and research agenda follow from the combined evidence?RQ5. 基于综合证据，应得出什么样的能力模型、课程路线图和研究议程？

The main contributions of this paper are fivefold. First, we offer a conceptual framework for AI-native software engineering organized around intent, collaboration, and verification. Second, we propose a multi-dimensional competency model that captures the skills needed for effective work in AI-mediated development environments. Third, we present a phased curriculum roadmap that supports educational transformation while preserving academic rigor and assessment validity. Fourth, we derive implications for faculty development and workforce preparation. Fifth, we identify a prioritized agenda of research gaps to guide future empirical and theoretical work. Together, these contributions aim to provide a rigorous foundation for rethinking software engineering for the AI-native era. The remainder of the paper proceeds as follows. Section 2 situates the work in the history of the discipline and prior reviews; Section 3 details the review methodology; Sections 4 and 5 report the scientometric and thematic findings; Section 7 critically interprets the contradictions in the evidence; Sections 8–11 develop the framework, competency model, curriculum roadmap, and faculty/workforce strategies; and Sections 12–14 present the research agenda with threats to validity, the conclusion, and data availability.本文的主要贡献有五个方面。首先，我们提供了一个围绕意图、协作和验证构建的AI原生软件工程概念框架。其次，我们提出了一个多维能力模型，涵盖了在AI中介开发环境中高效工作所需的技能。第三，我们提出了一个分阶段的课程路线图，在保持学术严谨性和评估有效性的同时支持教育转型。第四，我们推导了对教师发展和劳动力准备的影响。第五，我们确定了一个优先的研究空白议程，以指导未来的实证和理论工作。这些贡献旨在为重新思考AI原生时代的软件工程提供严谨的基础。本文其余部分安排如下：第2节将本研究置于学科历史和既往综述的背景下；第3节详述综述方法；第4节和第5节报告科学计量学和主题发现；第7节批判性地解读证据中的矛盾；第8–11节发展框架、能力模型、课程路线图以及教师/劳动力策略；第12–14节提出研究议程、效度威胁、结论及数据可用性。

Paradigm shifts in software engineering. The history of software engineering is marked by successive shifts in abstraction, tooling, and the division of labor between humans and machines. Assembly gave way to high-level languages; procedural code gave way to objects and components; waterfall planning gave way to agile and continuous delivery; and on-premise deployment gave way to cloud-native and platform engineering. Each transition raised the level of abstraction at which engineers expressed intent while preserving a deterministic relationship between what was written and what executed. Code-capable LLMs extend this trajectory, but they also introduce a qualitatively different interaction model: the primary interface becomes natural-language intent, and the resulting artifact is produced by a stochastic system rather than by a fully deterministic compiler or hand-authored procedure. Hassan et al. frame this transition as “SE 3.0,” an intent-first, conversation-driven paradigm in which humans collaborate with AI “teammates” [15]. In this view, the central challenge is not only code production, but also the translation, refinement, and verification of intent—a shift that relocates the locus of difficulty from syntax to specification and judgment.软件工程的范式转移。软件工程的历史以抽象、工具以及人机劳动分工的连续转变而著称。汇编语言让位于高级语言；过程式代码让位于对象和组件；瀑布式规划让位于敏捷和持续交付；本地部署让位于云原生和平台工程。每一次转变都提高了工程师表达意图的抽象水平，同时保持了所写内容与执行结果之间的确定性关系。具备代码能力的LLMs延续了这一轨迹，但它们也引入了一种质上不同的交互模型：主要接口变为自然语言意图，而最终产物由随机系统而非完全确定的编译器或手工编写的过程产生。Hassan等人将这一转变定义为“SE 3.0”，这是一种意图优先、对话驱动的范式，人类在其中与AI“队友”协作 [15]。在此视角下，核心挑战不仅是代码生产，还包括意图的转化、细化和验证——这一转变将难点从语法迁移到了规范说明和判断上。

Foundations of code LLMs. The current generation of code-oriented AI systems builds on a sequence of technical milestones. Early foundations include pre-trained representations such as CodeBERT [12] and CodeT5 [44], followed by the Codex model and the HumanEval benchmark [7], competition-level systems such as AlphaCode [25], and open, governance-aware models such as StarCoder [24]. These developments established the feasibility of code generation, completion, and synthesis across a range of programming tasks and progressively expanded the contexts, languages, and problem difficulties that models could address. At the same time, repository-level evaluations such as SWE-bench [20] showed that performance on realistic software engineering problems remains substantially more complex than performance on short, isolated programming tasks. This distinction is important because it highlights the gap between benchmark success on bounded functions and operational usefulness in real development settings, where issues span multiple files, require contextual understanding of a codebase, and demand changes that pass existing test suites.代码LLMs的基础。当前一代面向代码的AI系统建立在一系列技术里程碑之上。早期基础包括预训练表示，如CodeBERT [12] 和CodeT5 [44]，随后是Codex模型和HumanEval基准测试 [7]、竞赛级系统如AlphaCode [25]，以及开放且注重治理的模型如StarCoder [24]。这些进展确立了代码生成、补全和合成在各种编程任务中的可行性，并逐步扩展了模型可以处理的上下文、语言和问题难度。与此同时，诸如SWE-bench [20] 等仓库级别的评估显示，在现实软件工程问题上的表现仍然比在简短、孤立的编程任务上的表现复杂得多。这种区别很重要，因为它突显了在有界函数上的基准测试成功与在真实开发环境中的操作实用性之间的差距，在真实环境中，问题跨越多个文件，需要对代码库有上下文理解，并要求变更通过现有的测试套件。

Prior reviews and the gap addressed here. The literature has expanded rapidly, and several reviews have begun to organize this space. The most comprehensive synthesis to date, Hou et al. [19], reviews 395 LLM-for-SE studies and documents the scale and pace of this growth, mapping models to SE tasks, datasets, and open challenges. In parallel, computing-education syntheses, including the ITiCSE working-group report [36] and the Communications of the ACM review [11], examine how educational practice is responding to the same wave of technology. These reviews are valuable but tend to remain within a single literature—either the technical SE literature or the computing-education literature. Our review differs by integrating the practice, education, and workforce literatures in a single framework and by translating the combined evidence into an actionable competency model and curriculum roadmap. This integrative stance is what allows the contradictions observed in one literature (for example, context-dependent productivity effects) to inform recommendations in another (for example, how and when to teach verification and trust calibration).既往综述与本文解决的差距。文献扩展迅速，一些综述已开始整理这一领域。迄今为止最全面的综合研究Hou等人 [19] 回顾了395项LLM-for-SE研究，记录了这一增长的规模和速度，并将模型映射到SE任务、数据集和开放挑战中。与此同时，计算教育综合研究，包括ITiCSE工作组报告 [36] 和ACM通讯综述 [11]，考察了教育实践如何应对同一波技术浪潮。这些综述很有价值，但往往局限于单一文献——要么是技术SE文献，要么是计算教育文献。我们的综述不同之处在于将实践、教育和劳动力文献整合到一个单一框架中，并将综合证据转化为可操作的能力模型和课程路线图。这种整合立场使得在一个文献中观察到的矛盾（例如，情境依赖的生产力效应）能够为另一个文献（例如，如何以及何时教授验证和信任校准）中的建议提供参考。

We adopted a PRISMA-inspired review process executed through a multi-agent research workflow, with the aim of ensuring coverage across the main intellectual streams relevant to AI-native software engineering while preserving traceability, verification, and thematic balance. Four specialized search streams were used to interrogate complementary literatures and reduce the risk of narrow retrieval. Literature Discovery targeted code-capable LLMs, agentic software engineering, and SE-task automation. Scientometric Analysis focused on publication growth, venues, citation magnitudes, and signals of adoption. Curriculum Transformation examined computing-education venues and studies on pedagogical change. Workforce Impact addressed productivity, labor economics, and human–AI collaboration. Distributing discovery across four streams was a deliberate design choice: because the phenomenon under study spans technical, educational, organizational, and economic domains, a single search strategy anchored in one community would have systematically under-sampled the others.我们采用了受PRISMA启发并由多智能体研究工作流执行的综述流程，旨在确保涵盖与AI原生软件工程相关的主要知识流，同时保持可追溯性、验证性和主题平衡。使用了四个专门的搜索流来查询互补文献，并降低检索范围狭窄的风险。文献发现针对具备代码能力的LLMs、智能体软件工程和SE任务自动化。科学计量分析侧重于出版物增长、场所、引文规模和采用信号。课程转型考察计算教育场所和关于教学变革的研究。劳动力影响处理生产力、劳动经济学和人机协作。将发现分布在四个流中是一个刻意的设计选择：由于所研究的现象跨越技术、教育、组织和经济领域，锚定在一个社区的单一搜索策略会系统性地对其他领域采样不足。

Inclusion and exclusion criteria. We included peer-reviewed or high-credibility empirical works published between 2016 and 2026 that substantively addressed GenAI, LLMs, or agentic AI in relation to software engineering practice, education, or the workforce. To capture important adjacent evidence, we also included established-laboratory preprints and selected working papers in labor economics when they were credible and directly relevant, since several of the most consequential productivity and adoption results were first disseminated through these channels. We excluded records that were not verifiable, purely promotional material, and studies that were only peripheral to the review focus.纳入和排除标准。我们纳入了2016年至2026年间发表的、实质性探讨GenAI、LLMs或智能体AI与软件工程实践、教育或劳动力相关性的经同行评审或高可信度实证著作。为了捕捉重要的邻近证据，我们还纳入了已建立实验室的预印本和选定的劳动经济学工作论文（当它们可信且直接相关时），因为一些最具影响力的生产力和采用结果首先是通过这些渠道传播的。我们排除了不可验证的记录、纯粹的宣传材料以及仅与综述重点边缘相关的研究。

Screening and selection. Records identified by the four streams were screened for topical relevance and methodological credibility. When multiple streams surfaced the same study, duplicates were removed before final selection. The screening process was intentionally iterative: candidate studies were compared across themes, and the corpus was then refined to maintain balance across the three major trajectories of the review—namely practice, education, and workforce—so that no single trajectory dominated the synthesis simply because its literature was larger or faster-growing. From the broader evidence pool, we retained the 48 most influential and representative studies for the final corpus.筛选和选择。四个流识别出的记录被筛选以确定主题相关性和方法论可信度。当多个流出现同一研究时，在最终选择前删除了重复项。筛选过程是有意迭代的：候选研究在主题间进行比较，随后对语料库进行精炼以保持综述三个主要轨迹（即实践、教育和劳动力）之间的平衡，从而使单一轨迹不会仅仅因为其文献规模更大或增长更快而主导综合分析。从更广泛的证据池中，我们保留了48项最具影响力和代表性的研究作为最终语料库。

Verification. Every retained record was checked against its primary source whenever possible, using authoritative repositories and publishers including arXiv, ACL Anthology, ACM Digital Library, IEEE Xplore, publisher DOI pages, NBER, Science, and Management Science. A per-record confidence flag was assigned in the companion dataset to document verification status and source reliability. This step was used to improve transparency and to distinguish fully verified records from those requiring more cautious interpretation, which matters in a field where a large share of output circulates as preprints before formal peer review.验证。每份保留的记录尽可能地与原始来源进行了核对，使用了权威的存储库和出版商，包括arXiv、ACL Anthology、ACM Digital Library、IEEE Xplore、出版商DOI页面、NBER、Science和Management Science。在配套数据集中为每条记录分配了一个置信度标记，以记录验证状态和来源可靠性。此步骤用于提高透明度，并将完全验证的记录与需要更谨慎解读的记录区分开来，这在一个很大比例的产出在正式同行评审前以预印本形式流通的领域中非常重要。

The final corpus was synthesized across nine themes and mapped to the review questions through Table 1, while the complete study list is provided in Table LABEL:tab:corpus. The resulting dataset supports the PRISMA-style flow reported in Fig. 1 and underpins the conceptual, curricular, and workforce analyses presented in the remainder of the paper.最终语料库在九个主题上进行了综合，并通过表1映射到综述问题，而完整的研究列表在表LABEL:tab:corpus中提供。由此产生的数据集支持了图1中报告的PRISMA风格流程，并支撑了本文其余部分提出的概念、课程和劳动力分析。

The scientometric profile of the corpus reveals a pronounced acceleration in research activity following the emergence of publicly accessible generative AI systems. Consistent with the field-wide analysis reported by Hou et al. [19], the number of primary studies examining LLMs in software engineering increased from 7 in 2020 to 13 in 2021, before rising sharply to 56 in 2022 and 273 in 2023. This approximately five-fold increase coincides with the public release and widespread adoption of ChatGPT, marking a clear inflection point in both research attention and practical deployment (Fig. 2). The growth pattern suggests a rapid transition from exploratory investigations to large-scale scholarly engagement with AI-assisted software development, and it frames the temporal boundary that separates the pre- and post-ChatGPT phases of the literature analyzed here.语料库的科学计量概况揭示了在可公开访问的生成式AI系统出现后，研究活动出现了明显的加速。与Hou等人 [19] 报告的领域范围分析一致，检查LLMs在软件工程中应用的主要研究数量从2020年的7项增加到2021年的13项，随后在2022年急剧上升至56项，在2023年达到273项。这种约五倍的增长与ChatGPT的公开发布和广泛采用相吻合，标志着研究关注点和实际部署的一个明确转折点（图2）。增长模式表明，从探索性调查到对AI辅助软件开发的大规模学术参与发生了快速转变，并构成了本文分析的文献中ChatGPT前和后阶段的时间边界。

A similar trajectory is evident in computing-education research. The ITiCSE Working Group review identified 71 publications addressing generative AI in computing education, with approximately 80% appearing during the first eight months of 2023 alone [36]. This concentration of publications highlights the speed with which educational institutions and researchers responded to the pedagogical implications of generative AI technologies. Collectively, these trends indicate that software engineering and computing education have evolved in parallel, with advances in AI capability driving simultaneous changes in professional practice and instructional design rather than education lagging practice by the years that earlier technological shifts often exhibited.在计算教育研究中也存在类似的轨迹。ITiCSE工作组综述识别了71篇探讨生成式AI在计算教育中应用的出版物，其中约80%仅出现在2023年的前八个月 [36]。这种出版物的集中度突显了教育机构和研究人员对生成式AI技术教学影响的响应速度。总的来说，这些趋势表明软件工程和计算教育是平行演进的，AI能力的进步推动了专业实践和教学设计的同步变革，而不是像早期技术转变中经常表现出的教育滞后于实践数年的情况。

Evidence from industry surveys further suggests that adoption has progressed alongside capability improvements. Reported usage of AI-assisted development tools increased from approximately 76% of developers in 2024 to 84% in 2025. However, this increase in adoption was accompanied by declining levels of self-reported trust in AI-generated outputs. The resulting adoption–trust divergence is noteworthy because it implies that widespread tool utilization does not necessarily correspond to increased confidence in model reliability. This finding reinforces the importance of verification, critical evaluation, and human oversight as core competencies in AI-native software engineering, and it anticipates the “trust paradox” developed in the critical discussion.来自行业调查的证据进一步表明，采用率随着能力提升而同步进展。AI辅助开发工具的报告使用率从2024年的约76%增加到2025年的84%。然而，这种采用率的增加伴随着对AI生成输出的自我报告信任度下降。由此产生的采用-信任分歧值得注意，因为它意味着广泛的工具利用并不一定对应于对模型可靠性信心的增加。这一发现强化了验证、批判性评估和人类监督作为AI原生软件工程核心能力的重要性，并预示了批判性讨论中发展的“信任悖论”。

As illustrated in Fig. 3, the reviewed corpus exhibits substantial methodological diversity. Studies include tool and model evaluations, controlled and field experiments, qualitative and human–computer interaction investigations, surveys, benchmark studies, and position or vision papers. Such diversity reflects the interdisciplinary nature of the field and the absence of a single dominant methodological paradigm; it also means that synthesizing the evidence requires weighing findings produced under very different epistemic standards, from controlled randomized trials to interpretive qualitative accounts.如图3所示，所综述的语料库表现出实质性的方法论多样性。研究包括工具和模型评估、对照和实地实验、定性和人机交互调查、调查、基准测试研究以及立场或愿景论文。这种多样性反映了该领域的跨学科性质以及缺乏单一主导的方法论范式；这也意味着综合证据需要权衡在非常不同的认识论标准下产生的结果，从对照随机试验到解释性定性叙述。

The literature is distributed across leading venues in multiple disciplines. Software engineering contributions appear in ICSE, FSE, ASE, TSE, and TOSEM; machine-learning research is represented by ICLR, NeurIPS, ACL, and EMNLP; human–computer interaction studies appear in CHI, OOPSLA, and TOCHI; security-related work is published in S&P and CCS; educational research is drawn from SIGCSE, ITiCSE, ICER, ACE, and Koli Calling; and workforce-oriented evidence includes contributions from Science, Management Science, and NBER. This venue diversity underscores the broad impact of generative AI across technical, educational, organizational, and economic domains, and it is one reason a multi-stream discovery strategy was necessary to assemble a balanced corpus.文献分布在多个学科的领先场所。软件工程贡献出现在ICSE、FSE、ASE、TSE和TOSEM；机器学习研究由ICLR、NeurIPS、ACL和EMNLP代表；人机交互研究出现在CHI、OOPSLA和TOCHI；安全相关工作发表在S&P和CCS；教育研究来自SIGCSE、ITiCSE、ICER、ACE和Koli Calling；劳动力导向的证据包括来自Science、Management Science和NBER的贡献。这种场所多样性突显了生成式AI在技术、教育、组织和经济领域的广泛影响，这也是组建平衡语料库需要多流发现策略的原因之一。

A persistent characteristic of the field is the high proportion of preprints. More than half of the identified LLM-for-SE publications were disseminated initially through preprint channels before formal peer review. While this pattern reflects the exceptional pace of innovation and knowledge diffusion, it also introduces challenges related to replication, validation, and evidence stability. These concerns are revisited in the research-gap agenda and motivate continued emphasis on reproducibility and longitudinal evaluation, particularly where strong claims about productivity or learning effects rest on single studies that have not yet been independently replicated.该领域的一个持续特征是预印本比例很高。超过一半的已识别LLM-for-SE出版物在正式同行评审前最初是通过预印本渠道传播的。虽然这种模式反映了创新和知识传播的非凡速度，但它也引入了与复制、验证和证据稳定性相关的挑战。这些担忧在研究空白议程中被重新审视，并激发了对可重复性和纵向评估的持续强调，特别是在关于生产力或学习效果的强有力主张基于尚未被独立复制的单一研究时。

Table 2 summarizes the thematic distribution of the 48 studies across nine research themes, and Table LABEL:tab:corpus provides the full corpus with per-study venue, method, and core contribution. To convert this thematic catalogue into an interpretable account, we synthesize the evidence under three trajectories that cut across the themes: the advancing capability of the technology itself (practice), its contested effects on productivity and quality, and the response of education and the workforce.表2总结了48项研究在九个研究主题中的主题分布，表LABEL:tab:corpus提供了包含每项研究的场所、方法和核心贡献的完整语料库。为了将这一主题目录转换为可解释的叙述，我们在横跨这些主题的三个轨迹下综合了证据：技术自身能力的提升（实践）、其对生产力和质量的争议性影响，以及教育和劳动力的响应。

Three tensions structure the evidence. First, a productivity paradox: aggregate gains coexist with task-level slowdowns and hidden verification costs, so naive “X% faster” claims are misleading without controlling for expertise and context [32, 8, 4, 29]. The same technology that lets a novice complete a bounded task far faster can slow an expert working on a mature, high-stakes codebase, because the marginal value of a suggestion depends on how expensive it is to verify relative to writing the code directly. Second, a competence paradox: the same tools that lift novices’ immediate output may undermine the deliberate practice through which durable expertise forms, risking an “illusion of competence” in which fluent-looking results mask shallow understanding [38, 33]. Third, a trust paradox: adoption rises even as trust falls and measured security worsens [33, 31], making calibrated trust—knowing when to rely on and when to scrutinize AI output—rather than blanket acceptance or blanket rejection the pivotal skill. These tensions are not anomalies to be averaged away; they are structural features of human–AI collaboration that any educational or organizational response must confront directly. They converge on a single educational implication: the scarce, teachable human capability is no longer code production but judgment—specifying intent precisely, evaluating AI output critically, and verifying outcomes responsibly. This conclusion motivates the conceptual framework developed next.三种张力构成了证据结构。首先，生产力悖论：总体收益与任务级减速和隐藏验证成本并存，因此若不控制专业知识和上下文，幼稚的“X%更快”主张具有误导性 [32, 8, 4, 29]。同一项技术可以让新手在有界任务上完成得快得多，却可能拖慢在成熟、高风险代码库上工作的专家，因为建议的边际价值取决于相对于直接编写代码而言验证它的成本有多高。其次，能力悖论：提升新手即时产出的工具可能会破坏形成持久专业知识的刻意练习，冒着一种“能力错觉”的风险，即流利的外表掩盖了浅薄的理解 [38, 33]。第三，信任悖论：尽管信任下降且测得的安全性变差，采用率仍在上升 [33, 31]，使得校准信任——知道何时依赖以及何时审查AI输出——而不是盲目接受或盲目拒绝，成为关键技能。这些张力不是可以平均掉的异常；它们是人机协作的结构性特征，任何教育或组织响应都必须直接面对。它们汇聚成一个单一的教育启示：稀缺、可教授的人类能力不再是代码生产，而是判断——精确指定意图、批判性评估AI输出并负责任地验证结果。这一结论激发了接下来发展的概念框架。

Three tensions structure the evidence. First, a productivity paradox: aggregate gains coexist with task-level slowdowns and hidden verification costs, so naive “X% faster” claims are misleading without controlling for expertise and context [32, 8, 4, 29]. The same technology that lets a novice complete a bounded task far faster can slow an expert working on a mature, high-stakes codebase, because the marginal value of a suggestion depends on how expensive it is to verify relative to writing the code directly. Second, a competence paradox: the same tools that lift novices’ immediate output may undermine the deliberate practice through which durable expertise forms, risking an “illusion of competence” in which fluent-looking results mask shallow understanding [38, 33]. Third, a trust paradox: adoption rises even as trust falls and measured security worsens [33, 31], making calibrated trust—knowing when to rely on and when to scrutinize AI output—rather than blanket acceptance or blanket rejection the pivotal skill. These tensions are not anomalies to be averaged away; they are structural features of human–AI collaboration that any educational or organizational response must confront directly. They converge on a single educational implication: the scarce, teachable human capability is no longer code production but judgment—specifying intent precisely, evaluating AI output critically, and verifying outcomes responsibly. This conclusion motivates the conceptual framework developed next.

A further implication is that the corpus should not be read as evidence for linear substitution, but rather as evidence for reallocation of effort. The reviewed studies suggest that AI changes where cognitive load sits in the development process: less time may be spent on first-draft code, yet more time is often required for prompt refinement, output checking, debugging, and security review. In this sense, productivity gains are contingent on whether the surrounding workflow is designed to absorb the new verification burden efficiently. Where teams already possess strong systems knowledge, established testing practices, and disciplined review habits, AI can compress low-value work and expand higher-value design and integration work; where those supports are weak, the same tools can increase rework and obscure defects. This is why the literature does not support a universal productivity narrative. It instead indicates that AI magnifies existing process quality, rather than replacing it.进一步的含义是，语料库不应被解读为线性替代的证据，而应被解读为努力重新分配的证据。所综述的研究表明，AI改变了开发过程中认知负荷的位置：花费在初稿代码上的时间可能减少，但通常需要更多时间进行提示细化、输出检查、调试和安全审查。从这个意义上讲，生产力收益取决于周围的工作流是否设计为有效地吸收新的验证负担。在团队已经拥有强大的系统知识、既定的测试实践和严格的审查习惯的地方，AI可以压缩低价值工作并扩展更高价值的设计和集成工作；在这些支持薄弱的地方，同样的工具可能会增加返工并掩盖缺陷。这就是为什么文献不支持普遍的生产力叙事。它反而表明，AI放大了现有的流程质量，而不是取代它。

The educational corollary is equally important. If students are allowed to externalize too much early cognitive work, then apparent fluency may rise while conceptual retention weakens, creating a misleading signal of mastery. The competence paradox is therefore not simply about cheating or shortcutting, but about the conditions under which learning is still effortful enough to build robust mental models. In that respect, the emerging consensus is not anti-AI; it is pro-judgment. Educational and professional systems should cultivate engineers who can supervise AI responsibly, preserve conceptual depth, and decide when the machine should lead, assist, or be set aside.教育推论同样重要。如果允许学生外包过多的早期认知工作，那么表面的流利度可能会上升，而概念保留会减弱，从而产生掌握程度的误导信号。因此，能力悖论不仅仅是关于作弊或走捷径，而是关于学习在何种条件下仍然足够费力以建立稳健的心理模型。在这方面，新兴的共识不是反AI；而是支持判断。教育和专业系统应培养能够负责任地监督AI、保持概念深度并决定机器何时应主导、辅助或被搁置的工程师。

We organize the synthesis into a framework with three interacting pillars—Intent, Collaboration, and Verification—resting on a foundation of durable computer-science fundamentals and bounded by an ethics-and-security envelope (Fig. 4). Intent captures the upward migration to specification and prompt engineering: the engineer’s primary act becomes expressing what is wanted precisely enough that a stochastic system can act on it [10, 15]. Collaboration captures human–AI and human–agent teaming and orchestration, including the composition and supervision of multiple agents that play distinct software roles [2, 39, 47]. Verification captures critical evaluation, testing, security review, and trust calibration—the disciplined scrutiny that converts plausible output into trustworthy software [33, 41, 29]. The three pillars are mutually reinforcing: weak intent produces output that is harder to verify, weak verification makes collaboration unsafe, and weak foundations undermine all three by leaving the engineer unable to supervise the system. The framework rests on durable CS foundations because effective oversight of AI output presupposes understanding of algorithms, data structures, systems, and architecture; and it is bounded by an ethics, security, and responsible-use envelope because the documented security and equity risks are not optional add-ons but constraints on every pillar. The framework is the direct source of the competency model (Section 9) and the curriculum roadmap (Section 10).我们将综合内容组织成一个框架，具有三个相互作用的支柱——意图、协作和验证——建立在持久的计算机科学基础之上，并由伦理与安全包络所界定（图4）。意图捕获了向规范说明和提示工程的向上迁移：工程师的主要行为变为表达想要的内容，其精确程度足以让随机系统据此行动 [10, 15]。协作捕获人机和人-智能体团队协作与编排，包括组成和监督扮演不同软件角色的多个智能体 [2, 39, 47]。验证捕获批判性评估、测试、安全审查和信任校准——将合理输出转换为可信软件的严格审查 [33, 41, 29]。这三个支柱相互加强：薄弱的意图产生更难验证的输出，薄弱的验证使协作不安全，薄弱的基础通过使工程师无法监督系统而削弱所有三者。该框架建立在持久的CS基础之上，因为对AI输出的有效监督预设了对算法、数据结构、系统和架构的理解；它由伦理、安全和负责任使用包络所界定，因为记录在案的安全和公平风险不是可选的附加项，而是每个支柱的约束。该框架是能力模型（第9节）和课程路线图（第10节）的直接来源。

The conceptual value of the framework is that it provides a way to interpret apparently disparate findings within a single structure. Studies of prompting, conversational coding, debugging assistance, agentic orchestration, and security failures all become legible as evidence about different points on the same workflow. The framework makes explicit that AI-native software engineering is not defined by the presence of AI in the toolchain alone, but by a reorganization of responsibility across human and machine actors. Under this interpretation, the engineer does not vanish; instead, the engineer moves upward in abstraction, taking responsibility for defining goals, constraining behavior, and checking outcomes. This shift explains why the literature repeatedly returns to specification, verification, and supervision as central themes, even when the immediate use case appears to be code generation.该框架的概念价值在于，它提供了一种在单一结构内解释明显不同的发现的方法。关于提示、对话式编码、调试辅助、智能体编排和安全失败的研究，都可以作为关于同一工作流中不同点的证据而变得可读。该框架明确指出，AI原生软件工程不仅由工具链中AI的存在来定义，而且由人类和机器参与者之间责任的重组来定义。在这种解释下，工程师并没有消失；相反，工程师在抽象层面上向上移动，承担定义目标、约束行为和检查结果的责任。这种转变解释了为什么文献反复回到规范说明、验证和监督作为核心主题，即使即时用例看起来是代码生成。

The framework also clarifies why durable CS foundations remain indispensable. A supervisor cannot meaningfully evaluate output from a system whose logic, constraints, or failure modes are not understood at least at a working level. Likewise, collaboration with agents becomes unsafe when the human operator lacks the conceptual vocabulary to judge when to intervene. The ethics and security envelope is not a separate layer added after the fact; it is the condition under which all three pillars are acceptable in practice. Taken together, these relations define AI-native SE as a discipline of mediated creation: humans shape intention, coordinate systems of assistance, and retain responsibility for correctness, safety, and accountability.该框架还阐明了为什么持久的CS基础仍然不可或缺。监督者无法有意义地评估其逻辑、约束或失败模式至少在工作水平上未被理解的系统的输出。同样，当人类操作员缺乏判断何时干预的概念词汇时，与智能体的协作变得不安全。伦理和安全包络不是事后添加的独立层；它是所有三个支柱在实践中可接受的条件。总而言之，这些关系定义了AI原生SE作为一种中介创造的学科：人类塑造意图，协调辅助系统，并保留对正确性、安全性和问责制的责任。

Table 4 operationalizes the framework as nine competencies, each mapped to a dominant cognitive level and grounded in corpus evidence. The model deliberately elevates higher-order capabilities—specification, evaluation, orchestration, and metacognition—while retaining foundational CS knowledge as the basis for effective supervision of AI systems. The cognitive-level mapping, expressed in the vocabulary of a revised Bloom’s taxonomy, makes explicit why the model is weighted toward Evaluate and Create: in an environment where generation is cheap, the differentiating human contributions are those that judge, integrate, and direct. Each competency is traceable to specific evidence in the corpus, so the model is not an aspirational wish-list but a synthesis of what the reviewed studies indicate actually distinguishes effective from ineffective work with AI tools. The competencies are intended to be assessable and teachable, and they map onto the curriculum phases in Section 10.表4将框架操作化为九种能力，每种能力映射到一个主导的认知水平，并基于语料库证据。该模型刻意提升了更高阶的能力——规范说明、评估、编排和元认知——同时保留了基础CS知识作为有效监督AI系统的基础。以修订后的布鲁姆分类法词汇表达的认知水平映射，明确了为什么该模型偏向于评估和创造：在生成成本低廉的环境中，差异化的人类贡献是那些判断、整合和指导的贡献。每种能力都可以追溯到语料库中的具体证据，因此该模型不是一个抱负性的愿望清单，而是对所综述研究表明实际上将有效工作与无效工作与AI工具区分开来的综合。这些能力旨在可评估和可教授，并映射到第10节的课程阶段。

The nine competencies also help distinguish surface proficiency from robust capability. Competency C1, for example, is not simply about writing prompts, but about translating problem frames into precise specifications that constrain model behavior. C2 and C3 jointly capture the fact that AI output is only useful when coupled with critical reading, debugging, and verification practices. C4 recognizes that learners and practitioners need reflective habits that prevent overreliance on fluent but incorrect output. C5 reflects the emergence of orchestration as a meaningful engineering skill, particularly as workflows involve multiple tools or agents that must be coordinated rather than used in isolation. C6 serves as the enabling substrate, ensuring that the engineer can reason about the system rather than only interact with it. C7 foregrounds the security and ethical dimension, which is not reducible to compliance but tied directly to safe use. C8 and C9 complete the model by emphasizing communication, collaboration, and adaptability as enduring professional capacities.这九种能力还有助于区分表面熟练度与稳健能力。例如，能力C1不仅是关于编写提示，而是关于将问题框架转化为约束模型行为的精确规范说明。C2和C3共同捕获了这样一个事实：AI输出只有在与批判性阅读、调试和验证实践相结合时才有用。C4认识到学习者和从业者需要防止过度依赖流利但不正确输出的反思习惯。C5反映了编排作为一种有意义的工程技能的出现，特别是在工作流涉及必须协调而不是孤立使用的多个工具或智能体时。C6作为赋能基质，确保工程师能够对系统进行推理，而不仅仅是与之交互。C7突出了安全和伦理维度，它不能简化为合规，而是直接与安全使用相关。C8和C9通过强调沟通、协作和适应性作为持久的专业能力来完成模型。

Taken together, these competencies define an assessment logic for the AI-native era. The point is not to eliminate traditional measures of correctness, but to supplement them with assessments that reveal whether students or practitioners can inspect output, defend decisions, and operate responsibly in uncertain conditions. This is especially important because fluent interaction with AI can mask weaknesses in explanation, reasoning, or verification. A competency model therefore has value beyond curricular mapping: it gives institutions a language for diagnosing gaps, designing interventions, and tracking progress as AI capabilities continue to evolve.总而言之，这些能力定义了AI原生时代的评估逻辑。重点不是消除传统的正确性衡量标准，而是用揭示学生或从业者是否能够检查输出、捍卫决策并在不确定条件下负责任地操作的评估来补充它们。这尤为重要，因为与AI的流利交互可以掩盖解释、推理或验证方面的弱点。因此，能力模型具有超越课程映射的价值：它为机构提供了一种语言，用于在AI能力持续演进时诊断差距、设计干预措施和跟踪进展。

We propose a four-phase integration model (Table 5) that protects deliberate practice early, then progressively shifts toward human–AI teaming and authentic, agentic projects. The unifying principle is assessment realignment: because code-writing tasks are now AI-solvable [13, 14], assessment must privilege process, specification, evaluation, and defense of work over artifact production alone. The phasing is itself a response to the competence paradox: early courses deliberately restrict AI on core skill-building so that students develop the mental models that later make them effective supervisors of AI, while later courses progressively open up to AI collaboration as the assessment focus shifts from producing artifacts to directing, evaluating, and defending them. Across all phases, a cross-cutting strand threads responsible use, security, intellectual property, equity, and integrity through the curriculum rather than isolating them in a single ethics course, reflecting the framework’s ethics-and-security envelope.我们提出了一个四阶段集成模型（表 5），该模型在早期阶段保护刻意练习，随后逐步转向人机协作以及真实的、具代理能力的（agentic）项目。其核心原则是评估重构：由于代码编写任务现在已能由人工智能解决 [13, 14]，评估必须优先考虑过程、规范、评估和作品辩护，而不仅仅是产出制品。这种分阶段本身是对“能力悖论”的回应：早期课程刻意限制人工智能在核心技能培养中的使用，以便学生构建出心智模型，从而在后续阶段能够有效地监督人工智能；而后期课程则随着评估重点从制品产出转向指导、评估和辩护，逐步开放人工智能协作。在所有阶段中，课程体系贯穿了负责任使用、安全、知识产权、公平性和诚信等跨学科主线，而非将其孤立在单一的伦理课程中，这体现了该框架的伦理与安全包络（ethics-and-security envelope）设计。

The roadmap is best understood as a gradual transformation of academic judgment rather than a simple increase in tool access. In the foundational phase, the aim is to ensure that students can still reason, trace, and explain computational behavior without relying on generative systems. This is not a retreat from AI, but a necessary condition for later effective use. As learners advance, AI becomes a subject of analysis and a collaborator in bounded tasks, allowing students to experience both the strengths and limitations of model-assisted development. By the time learners reach systems, SE, and capstone contexts, they should be ready to manage larger workflows in which the central challenge is no longer whether AI can produce code, but whether the human team can specify, integrate, audit, and defend what the AI produces.该路线图最好被理解为学术判断力的逐步转型，而非仅仅是工具获取能力的简单提升。在基础阶段，目标是确保学生在不依赖生成式系统的情况下，依然能够进行推理、追踪和解释计算行为。这并非对人工智能的退缩，而是后续有效使用的必要条件。随着学习者的进步，人工智能将成为分析的对象以及有界任务中的协作伙伴，使学生能够同时体验到模型辅助开发的优势与局限。当学习者进入系统、软件工程（SE）和毕业设计阶段时，他们应已准备好管理更大型的工作流，此时的核心挑战不再是人工智能能否生成代码，而是人类团队能否对人工智能的产出进行规范、集成、审计和辩护。

The roadmap also has institutional significance. It implies that curriculum reform cannot be confined to isolated elective modules or brief policy statements. Programs need a coherent sequence in which learning outcomes, assessment designs, and academic integrity practices reinforce one another. That sequence should be backed by faculty development and shared assessment resources, because otherwise AI-resilient design will remain uneven and difficult to sustain. The roadmap therefore serves both as a pedagogical model and as a governance mechanism for curriculum modernization.该路线图也具有机构层面的意义。它意味着课程改革不能局限于孤立的选修模块或简短的政策声明。各专业需要一个连贯的序列，使学习成果、评估设计和学术诚信实践能够相互强化。该序列应得到教师发展和共享评估资源的支持，否则抗人工智能（AI-resilient）的设计将难以均匀分布且难以持续。因此，该路线图既是一种教学模型，也是课程现代化的治理机制。

Faculty development. Because instructor responses currently range from prohibition to wholesale integration [22], institutions need structured, evaluated faculty-development programs rather than ad hoc adaptation. We recommend: (i) communities of practice that co-design AI-resilient assessments so that effective designs spread rather than being reinvented course by course; (ii) hands-on training in agentic and tutoring tools with guardrails, building on evidence about where such tools help and where they remain unreliable [27, 35]; (iii) shared, openly-licensed assessment banks that lower the cost of moving away from easily-automated artifact production; and (iv) longitudinal evaluation of what faculty actually implement, addressing the well-documented intention–action gap in which stated plans to integrate or restrict AI diverge from classroom practice.教师发展。由于目前讲师的反应从全面禁止到全面集成各不相同 [22]，各机构需要结构化、可评估的教师发展计划，而非临时的适应性调整。我们建议：(i) 建立实践社区，共同设计抗人工智能的评估方式，使有效的设计得以传播，而不是在每门课程中重复发明；(ii) 对具备防护措施的代理工具和辅导工具进行实操培训，并基于证据明确此类工具在何处有用、在何处仍不可靠 [27, 35]；(iii) 建立共享且开放授权的评估题库，降低放弃易于自动化的制品产出评估的成本；(iv) 对教师的实际实施情况进行纵向评估，解决“意图与行为脱节”这一有据可查的问题，即教师计划集成或限制人工智能的意图与实际课堂教学实践存在偏差。

Faculty development should be treated as a design problem, not only a training problem. The central challenge is not merely that instructors need to learn new tools, but that they need new assessment norms, new examples of acceptable use, and new ways of judging student learning when AI is present. Many faculty are likely to benefit from concrete templates that make it easier to ask for evidence of reasoning, reflection, and verification. Others may need support in identifying where AI assistance is pedagogically useful and where it undermines the learning goals of a particular course. A mature faculty-development program therefore combines technical exposure, pedagogical calibration, and policy guidance. It should also create space for instructors to share what works and what fails, because the field is changing too quickly for isolated experimentation to scale reliably.教师发展应被视为一个设计问题，而不仅仅是培训问题。核心挑战不仅在于讲师需要学习新工具，更在于当人工智能存在时，他们需要新的评估准则、新的可接受使用范例，以及评判学生学习成果的新方法。许多教师可能会从具体的模板中受益，这些模板能更轻松地要求学生提供推理、反思和验证的证据。其他人可能需要支持，以识别在哪些地方人工智能辅助在教学上是有用的，在哪些地方它会削弱特定课程的学习目标。因此，一个成熟的教师发展计划应结合技术接触、教学校准和政策指导。它还应为讲师创造分享成功与失败经验的空间，因为该领域发展太快，孤立的实验难以实现可靠的规模化。

Workforce transformation. Evidence that AI most benefits novices [8, 6] yet can slow experts on mature systems [4] implies differentiated reskilling rather than a single organization-wide policy. Junior pipelines should emphasize verification and systems understanding to avoid an “illusion of competence,” ensuring that early-career engineers build the judgment needed to supervise AI rather than merely accept its output; senior staff, by contrast, need patterns for when not to delegate, recognizing the contexts in which manual work remains faster and safer. As effort reallocates from coordination to core coding [17] and roles reorganize around intent and orchestration [15], organizations should invest in developer-experience measurement that explicitly accounts for verification overhead [29] rather than crediting raw acceptance or output, and in secure-by-default AI workflows given the documented security regressions [31, 33]. The through-line connecting faculty development and workforce transformation is that both are governance problems as much as training problems: the durable gains come from institutionalizing verification, calibrated trust, and responsible use, not from tool adoption alone.劳动力转型。证据表明，人工智能最有利于新手 [8, 6]，却可能拖慢专家在成熟系统上的工作 [4]，这意味着需要差异化的技能重塑，而非单一的组织范围政策。初级人才培养应强调验证和系统理解，以避免“能力错觉”，确保早期职业工程师建立起监督人工智能所需的判断力，而非仅仅接受其输出；相比之下，资深员工需要明确何时不应委派任务的模式，识别在哪些场景下手动工作依然更快、更安全。随着工作重点从协调转向核心编码 [17]，角色围绕意图和编排进行重组 [15]，组织应投资于明确考虑了验证开销 [29] 的开发者体验度量，而不是单纯认可原始的采纳率或产出量；同时，鉴于已记录的安全回归问题 [31, 33]，应投资于默认安全的 AI 工作流。连接教师发展与劳动力转型的核心在于，两者都是治理问题，而不仅仅是培训问题：持久的收益来自于制度化的验证、校准后的信任以及负责任的使用，而非仅仅是工具的采用。

At the workforce level, the evidence suggests a need for role-sensitive transformation. Entry-level engineers should be equipped to question, inspect, and verify AI-assisted work from the outset, because their value will increasingly lie in disciplined judgment rather than in isolated coding speed. More experienced engineers, meanwhile, must learn to recognize when AI assistance is appropriate, when it adds unnecessary verification burden, and when existing expertise remains the more efficient and safer route. For organizations, this means that talent development should not focus only on accelerating output. It should also aim to preserve organizational memory, strengthen review culture, and prevent overreliance on tools whose performance may vary sharply with task complexity and domain maturity. In that sense, workforce transformation is less about replacing expertise than about redefining what expertise must now include.在劳动力层面，证据表明需要进行角色敏感的转型。初级工程师应从一开始就具备质疑、检查和验证人工智能辅助工作的能力，因为他们的价值将日益体现在严谨的判断力而非单纯的编码速度上。同时，更有经验的工程师必须学会识别何时适合使用人工智能辅助，何时它增加了不必要的验证负担，以及何时现有的专业知识依然是更高效、更安全的途径。对于组织而言，这意味着人才发展不应仅关注加快产出，还应旨在保护组织记忆、加强评审文化，并防止过度依赖那些性能随任务复杂度和领域成熟度剧烈波动的工具。从这个意义上说，劳动力转型与其说是取代专业知识，不如说是重新定义专业知识现在必须包含的内容。

Future research agenda. Table 6 consolidates eleven gaps and directions synthesized across the corpus. The highest priorities are longitudinal learning and skill-formation studies that follow learners across multiple semesters rather than single sessions; quality- and team-adjusted productivity measurement that captures verification and maintenance costs rather than raw speed; a theory of when AI helps versus hinders, expressed in terms of moderators such as expertise, task novelty, and codebase maturity; validated AI-resilient assessment instruments; and equity-focused interventions targeted at the widening gap between strong and struggling learners. These priorities follow directly from the contradictions documented in the critical discussion: each gap marks a place where the existing evidence is either too short-term, too narrowly measured, or too concentrated in a few contexts to support confident generalization.未来研究议程。表 6 整合了从文献综述中归纳出的 11 个差距和方向。最优先的任务包括：跨越多个学期而非单次会话的纵向学习与技能形成研究；能够捕获验证与维护成本而非仅衡量原始速度的质量与团队调整后的生产力指标；关于人工智能何时助益或阻碍的理论，并以专业知识、任务新颖性和代码库成熟度等调节变量来表达；经过验证的抗人工智能评估工具；以及针对强弱学习者之间差距扩大的公平性干预措施。这些优先事项直接源于批判性讨论中记录的矛盾：每一个差距都标志着现有证据要么过于短期、衡量过于狭窄，要么过于集中在少数场景中，从而无法支持可靠的概括。

Threats to validity. Several limitations qualify the conclusions drawn here. On construct and selection validity, “influence” is partly subjective; we mitigated this with multi-agent discovery across complementary literatures and explicit verification, but some relevant work is inevitably omitted. On currency, the field moves faster than publication; over half of LLM-for-SE outputs are preprints [19], so some venues and citation magnitudes are reported as orders of magnitude and may have changed since data collection. On internal validity, many primary studies use bounded tasks or single institutions; we therefore foreground the contradictions in the evidence rather than averaging them away, since a pooled effect size would obscure the very context-dependence that is the central finding. On external validity, the education evidence concentrates on introductory programming and a small set of research groups, and output is geographically concentrated; generalization beyond these contexts is uncertain and is itself flagged as a research gap (G11). These threats do not undermine the review’s directional conclusions, but they bound the precision with which any single magnitude should be reported or acted upon.效度威胁。本文得出的结论受到若干局限性的制约。在构念效度和选择效度方面，“影响”具有一定的主观性；我们通过跨互补文献的多智能体发现和明确的验证来减轻这一影响，但不可避免地遗漏了一些相关工作。在时效性方面，该领域的发展速度超过了出版速度；超过一半的软件工程人工智能（LLM-for-SE）成果是预印本 [19]，因此某些场所和引用量级仅作为量级报告，在数据收集后可能已发生变化。在内部效度方面，许多主要研究使用有界任务或单一机构；因此，我们突出了证据中的矛盾，而不是将其平均化，因为汇总后的效应量会掩盖作为核心发现的上下文依赖性。在外部效度方面，教育证据集中在入门编程和少数研究小组，且产出在地理上高度集中；在这些场景之外的概括性尚不确定，这本身也被标记为研究差距（G11）。这些威胁并未削弱本文的指导性结论，但它们限制了任何单一量级在报告或应用时的精确度。

To maintain the same scholarly tone, synthesis-driven style, and conclusion-oriented focus while expanding the section to approximately 240 words, you can use the following version: 为了保持相同的学术基调、综合驱动的风格和以结论为导向的重点，同时将该部分扩展至约 240 字，您可以使用以下版本：

Across 48 verified studies, the evidence is consistent on direction if not magnitude: GenAI, LLMs, and emerging agentic systems are fundamentally reshaping software engineering by shifting human effort away from code authorship and toward intent specification, collaboration, supervision, and verification. Although the literature reports substantial variation in measured outcomes, a common pattern emerges. Productivity gains, learning benefits, and workflow improvements are achievable, but they are neither automatic nor universal. Instead, their realization depends on expertise, task characteristics, organizational context, and the quality of verification practices surrounding AI use.通过对 48 项已验证研究的分析，证据在方向上是一致的，尽管在量级上存在差异：生成式人工智能（GenAI）、大语言模型（LLMs）以及新兴的代理系统正通过将人类精力从代码编写转向意图规范、协作、监督和验证，从根本上重塑软件工程。尽管文献报告的测量结果存在显著差异，但一个共同的模式已经显现。生产力的提升、学习收益和工作流的改进是可以实现的，但它们既不是自动的，也不是普遍的。相反，这些成果的实现取决于专业知识、任务特征、组织背景以及围绕人工智能使用的验证实践的质量。

This review identified three recurring tensions that define the AI-native era: the productivity paradox, the competence paradox, and the trust paradox. Collectively, these tensions suggest that the central challenge is no longer generating software artifacts, but ensuring that humans retain the judgment necessary to direct, evaluate, and govern increasingly capable AI systems. In response, this paper synthesized the evidence into an integrated conceptual framework, a nine-dimension competency model, a four-phase curriculum roadmap with AI-resilient assessment, and a set of faculty-development and workforce-transformation recommendations.本综述确定了定义人工智能原生时代的三个反复出现的张力：生产力悖论、能力悖论和信任悖论。总的来说，这些张力表明核心挑战不再是生成软件制品，而是确保人类保留指导、评估和治理日益强大的人工智能系统所需的判断力。为此，本文将证据综合为一个集成的概念框架、一个九维度能力模型、一个包含抗人工智能评估的四阶段课程路线图，以及一套教师发展和劳动力转型的建议。

The findings carry important implications for universities, employers, and policymakers. Educational programs that preserve deliberate practice while emphasizing specification, evaluation, verification, and responsible use are likely to produce graduates better prepared for AI-mediated development environments. Likewise, organizations that institutionalize calibrated trust, secure workflows, and effective human oversight will be better positioned to capture the benefits of AI while mitigating its risks. Ultimately, preparing software engineers for the AI-native future is not primarily a technological challenge; it is an educational, organizational, and governance challenge centered on cultivating enduring human judgment in a rapidly evolving technological landscape.这些发现对于大学、雇主和政策制定者具有重要意义。那些在保持刻意练习的同时强调规范、评估、验证和负责任使用的教育项目，更有可能培养出为人工智能中介的开发环境做好准备的毕业生。同样，那些将校准后的信任、安全的工作流和有效的人类监督制度化的组织，将更有能力在利用人工智能收益的同时减轻其风险。归根结底，为人工智能原生的未来培养软件工程师，主要不是一个技术挑战；而是一个以在快速演变的技术环境中培养持久人类判断力为中心的教育、组织和治理挑战。

The collected data is shown within the manuscript in detail as a table. Readers may reach out to the author to obtain a companion spreadsheet that provides full bibliographic and analytical coding for all 48 studies, plus thematic, temporal, methodological, venue, gap, competency, curriculum, and PRISMA sheets.收集的数据已在文稿中以表格形式详细展示。读者可联系作者获取配套电子表格，其中提供了全部 48 项研究的完整书目和分析编码，以及主题、时间、方法、来源、差距、能力、课程和 PRISMA 工作表。