~/shanegraffiti.com/research/ai-native-se
Shane Graffiti Inc. Semantic Adversarial Research Division 2026

THE RISE OF
AI-NATIVE
SOFTWARE
ENGINEERING

Generative AI, large language models, and emerging agentic systems constitute the most disruptive transformation in the history of software engineering reshaping development processes, required competencies, professional roles, and the outcomes universities must deliver. A systematic review of 48 verified, influential peer-reviewed studies (2016–2026) synthesises the evidence along three trajectories: practice, education, and workforce. The corpus reveals a scientometric inflection annual LLM-for-SE output grew roughly five-fold after late 2022 and an evidence base internally contradictory on the size and even the sign of productivity effects. The central lesson: educating engineers for judgment, verification, and orchestration, rather than code production alone, is the defining challenge of the AI-native era.

Division Semantic Adversarial Research
Domain SE Practice / Education / Workforce
Published arXiv 2606.12986 2026
Corpus 48 studies · ~5× post-2022 growth
SE 3.0 Intent Specification Agent Orchestration Verification SWE-bench Calibrated Trust The Productivity Paradox The Competence Paradox The Trust Paradox AI-Resilient Assessment Human–AI Collaboration Curriculum Roadmap SE 3.0 Intent Specification Agent Orchestration Verification SWE-bench Calibrated Trust The Productivity Paradox The Competence Paradox The Trust Paradox AI-Resilient Assessment Human–AI Collaboration Curriculum Roadmap
§ 1.0 The Scientometric Inflection

Research activity accelerated sharply after publicly accessible generative AI arrived. Field-wide, the number of primary LLM-for-SE studies rose from 7 in 2020 to 13 in 2021, then jumped to 56 in 2022 and 273 in 2023 an inflection point that coincides with the late-2022 public release of ChatGPT and a parallel surge in computing-education research.

7
2020
13
2021
56
2022
273
2023

// primary LLM-for-SE studies per year     ~5× rise from 2022 to 2023
// adoption climbed 76% → 84% of developers (2024→2025), yet self-reported trust fell the adoption–trust divergence.

§ 2.0 Three Trajectories

The synthesis cuts across nine themes into three trajectories that move together. If entry-level coding becomes automatable, then what universities assess, what employers hire for, and what a "software engineer" means must all change at once.

Practice Completion to Autonomy
The capability frontier advanced from token-level completion to repository-scale autonomy in roughly three years. SWE-bench reframed evaluation around real GitHub issues; reasoning-and-acting and self-reflection paradigms enabled tool-using agents and multi-agent frameworks. Human effort migrated up the abstraction stack from writing statements to specifying intent and verifying outcomes.
Productivity A Contested Signal
Controlled trials report large gains a 55.8% speed-up on a bounded task, ~26% more completed tasks across three firm experiments with novices benefiting most. Yet other studies find no reliable improvement, large hidden verification costs, and experienced developers ~19% slower on mature codebases while believing themselves faster. Effect size, and even sign, depend on expertise and context.
Education & Workforce From Alarm to Redesign
Code-capable models outperform most students on CS1 and CS2 exams, forcing assessment to privilege process over artefact. The workforce reallocates effort toward core coding and away from coordination, reframing the engineer's identity around intent and orchestration. The scarce, valuable capabilities are framing problems, evaluating output, and integrating it responsibly.
§ 3.0 Three Structural Paradoxes

Three tensions structure the evidence. They are not anomalies to be averaged away they are structural features of human–AI collaboration that any educational or organisational response must confront directly.

Paradox 01
Productivity
Aggregate gains coexist with task-level slowdowns and hidden verification costs. The same tool that lets a novice finish a bounded task far faster can slow an expert on a mature, high-stakes codebase because a suggestion's value depends on how costly it is to verify relative to writing the code directly.
Paradox 02
Competence
The 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. The risk is not cheating; it is learning that is no longer effortful enough to build robust mental models.
Paradox 03
Trust
Adoption rises even as trust falls and measured security worsens roughly 40% of generated programs were vulnerable in security-sensitive scenarios, and assisted users wrote less secure code while feeling more confident. The pivotal skill is calibrated trust: knowing when to rely on output and when to scrutinise it.
§ 4.0 The AI-Native Framework

The synthesis organises into three mutually reinforcing pillars resting on durable computer-science foundations and bounded by an ethics-and-security envelope. Weak intent produces output that is harder to verify; weak verification makes collaboration unsafe; weak foundations leave the engineer unable to supervise the system at all.

Ethics, Security & Responsible-Use Envelope
Intent
Specification, prompt and problem engineering. The engineer's primary act becomes expressing what is wanted precisely enough that a stochastic system can act on it.
Collaboration
Human–AI and human–agent teaming. Composing and supervising multiple agents that play distinct software roles, rather than using tools in isolation.
Verification
Critical evaluation, testing, security review, trust calibration the disciplined scrutiny that converts plausible output into trustworthy software.
Durable CS Foundations
Algorithms · data structures · systems · architecture effective oversight presupposes understanding the system, not only interacting with it.
§ 5.0 Nine-Dimension Competency Model

The framework operationalises as nine competencies, each mapped to a dominant cognitive level. Where generation is cheap, the differentiating human contributions are those that judge, integrate, and direct so the model is weighted toward Evaluate and Create.

IDCompetencyCognitive level
C1Specification & intent engineeringCreate / Evaluate
C2Critical evaluation of AI outputEvaluate / Analyze
C3AI-assisted debugging & verificationApply / Analyze
C4Metacognition & self-regulationEvaluate
C5Agent orchestration & tool useCreate / Apply
C6Foundational CS & systems thinkingUnderstand / Apply
C7Security, ethics & responsible useApply / Evaluate
C8Human–AI collaboration & communicationApply / Create
C9Continuous learning & adaptabilityCreate
§ 6.0 Four-Phase Curriculum Roadmap

A phased model 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, assessment must privilege process, specification, evaluation, and the defence of work over artefact production alone.

P1
CS1 / CS2 Foundations
Durable fundamentals and AI literacy, restricting AI on core skill-building so students build the mental models that later make them effective supervisors.
Assessment: invigilated/oral fundamentals · code tracing · prompt-problem tasks
P2
Core Data Structures & Design
AI as a studied collaborator. The focus moves to design, testing, and verification in bounded tasks where strengths and limits are experienced firsthand.
Assessment: test-adequacy tasks · code-review portfolios · design rationales
P3
SE & Systems
Human–AI teams, agent orchestration, and quality and security at scale managing larger workflows rather than producing isolated code.
Assessment: team projects with AI teammates · defect/security metrics · reflective logs
P4
Capstone & Electives
Authentic, repo-scale, agentic projects with governance. The central challenge is no longer whether AI can produce code, but whether the human team can specify, integrate, audit, and defend it.
Assessment: public defence · contribution & process evidence
~5×
Research growth, 2022→2023
40%
Generated programs vulnerable in security tests
~19%
Experts slower on mature repos (believed faster)
11
Prioritised research gaps identified
~/conclusion
$ query: what is the evidence consistent on // direction, not magnitude. // effort shifts from authoring code to specifying, supervising, verifying. $ query: are the productivity gains real // yes but neither automatic nor universal. // they depend on expertise, task, context, and verification quality. $ query: what is the central challenge // not generating artifacts retaining the judgment to govern them. // an educational and governance problem, not a technological one.

THE
SCARCE
SKILL IS
JUDGMENT,
NOT
CODE.