Source-linked AI summary
Why Software Engineering Is Indispensable in the Age of Coding Agents
Alfonso Fuggetta
TL;DR
The paper asks whether AI coding agents make software engineering and software engineers obsolete, against a backdrop of emerging but mixed evidence. It explains agents’ structural knowledge gaps and argues that engineers and persistent artifacts remain necessary.
Problem
The paper asks whether capable AI coding agents make software engineering obsolete, while emerging evidence on their effects remains mixed across settings.
Source: §1 · p. 1Method
Results
The paper concludes that software engineering and software engineers are essential, not obsolete, to defining the structured space in which AI agents can be trusted.
Source: §6 · p. 5Takeaways & Limitations
Organizations should treat specifications, domain elicitation, architectural governance, and process discipline as the structured space for trustworthy AI-agent operation, not overhead to minimize.
Source: §6 · p. 5
Abstract
from arXiv · showhide
Can AI make Software Engineering (SE) -- the discipline -- obsolete? And can it make software engineers -- the professionals -- redundant? This paper argues that the rise of capable AI coding agents makes SE and software engineers essential, not obsolete: the missing foundation without which AI-assisted development produces misleadingly plausible, unverifiable, and ultimately untrustworthy software. Three structural properties of large language models (probabilistic generation, agnosticism, and semantic statelessness) create a structural vacuum that no amount of training can eliminate. Filling it requires four knowledge levers: methodological knowledge, domain knowledge, design choices, and process choices. All four must be reified as persistent artifacts, and each requires the software engineer as methodologist, mediator, and custodian.
1 What We See
Evidence on AI coding agents is mixed across tasks and settings, while professional practice relies on planning, validation, and architectural control. The section argues that software engineering supplies structure needed to address agents’ limits.
1 What We See
- Meyer argues vibe coding suits casual software but not acute or business software, where correctness is paramount, coding is only 10–20% of project cost, and formal specification and verification are needed.He names a “hallucination loop” as a plausible suggestion that sends the developer down a path that will not close.Source: §1 · p. 1
- Across studies, AI effects varied with task complexity, codebase maturity, and developer seniority, with gains on delimited work offset by supervision and rework on complex tasks.Cui et al. found a 26% increase in completed tasks across 4,867 developers; Becker et al. [2] found developers took 19% longer; Xu et al. [10] reported aggregate line-of-code output rose 17.7%, while core contributors’ original code output fell 19%.Source: §1 · p. 1
- Professionals plan, validate agent output, and retain architectural control, unlike developers who treat interaction flow as evidence of correctness.The field study included thirteen developers and a survey of ninety-nine more.Source: §1 · p. 2
- In a survey of nearly 49,000 developers across 177 countries, 84% adopted AI tools, 46% distrusted output accuracy, and 66% cited “almost right, but not quite” solutions as the top frustration.Distrust rose from 31%.Source: §1 · p. 2
- Software’s evolving, inconsistent requirements make it unlike a finished classical-engineering artifact; the paper argues that SE supplies what AI lacks, while AI can accelerate SE’s systematic application.The discipline has developed formal methods, safety standards, empirical software engineering, model-based systems engineering, and product line engineering.Source: §1 · p. 2
2 Why: The Structural Vacuum
The paper explains AI coding agents’ limits as a structural condition rather than a performance gap. Their properties leave irreducible knowledge and judgment needs that require human reasoning and external artifacts.
2 Why: The Structural Vacuum
- The paper uses a logical and epistemological analysis of agents’ structure to derive the roles human knowledge and judgment must play; it characterizes the vacuum as structural, not a performance gap.Source: §2 · p. 2
- Probabilistic generation means an LLM generates output conditioned on training and context, and no artifact can be validated against its represented world without an external reference.Source: §2 · p. 2
- Methodological and domain agnosticism have reducible surfaces, but evaluating a specification requires domain-grounded judgment, while organization-specific knowledge is continuously renegotiated and can only be managed.Specialized models can absorb SE principles, and general domain knowledge can be supplied through retrieval-augmented generation, fine-tuning, or domain training.Source: §2 · p. 2
- Evolution patterns may be trainable, but a system’s project-specific trajectory under organizational constraints does not exist until a software engineer reasons it into existence.Source: §2 · pp. 2–3
- Semantic statelessness leaves prior decisions and domain constraints unusable as accumulated know-how unless they are reified as artifacts, even when systems retain cross-session logs or storage.Source: §2 · p. 3
3 Four Knowledge Levers, Three Actors
Filling the structural vacuum requires four kinds of knowledge to be made persistent, with domain experts, AI agents, and software engineers playing distinct roles. The paper also argues these limits cannot be trained away.
3 Four Knowledge Levers, Three Actors
- The four levers are methodological knowledge, domain knowledge, project-specific design choices, and process choices.They respectively concern how to build quality software, what the system means and does in its world, project decisions such as architecture and interfaces, and how work is organized.Source: §3 · p. 3
- All four levers must be reified as persistent, accessible artifacts because knowledge held only in someone’s head is unavailable to a stateless agent.Source: §3 · p. 3
- Domain experts supply semantic knowledge, the AI agent executes without an intrinsic correctness reference or future orientation, and the software engineer mediates and maintains the system’s trajectory.Source: §3 · p. 3
- The irreducible cores—evaluative judgment, organizational knowledge, and a system’s future trajectory—cannot be distilled from any corpus, so the vacuum is structural rather than historical.Source: §3 · p. 3
4 SE and AI
SE fills the coding agent’s structural vacuum through methods embedded in the agent and project-specific artifacts supplied by engineers. AI can also lower the cost of applying established SE practices.
4 SE and AI
- SE fills the reducible layer through constitutive methods embedded in agent design and the irreducible core through operational project artifacts.Constitutive methods prompt for specifications and structure outputs; operational artifacts provide project-specific references for correctness.Source: §4 · p. 3
- AI lowers the cost of reification, verification, and accountability, making specifications, tests, and Architecture Decision Records practical substrates for agent work.Specifications can emerge from elicitation, tests can become the agent’s reference, and Architecture Decision Records can guide its work; Shaw and colleagues call the resulting skill design rather than coding.Source: §4 · pp. 3–4
- Software engineers act methodologically, by structuring SE; mediatively, by eliciting domain knowledge; and custodially, by owning design and process choices.The DORA 2025 State of AI report finds that AI adoption amplifies existing practices: mature foundations convert AI into delivery throughput, while teams without them find it accelerates instability rather than productivity.Source: §4 · p. 4
5 In Practice: Artifacts and Workflows
The paper describes persistent artifacts that support agent work across software lifecycles, while arguing that AI can support workflows without prescribing one. Its probabilistic operation also opens a direction for process governance beyond deterministic automation’s narrow technical phases.
5 In Practice: Artifacts and Workflows
- The artifact list is non-exhaustive and largely independent of the software lifecycle adopted.AI can support waterfall, agile, spec-driven, or incremental refinement workflows if artifacts are present and continuously maintained.Source: §5 · p. 4
- Specifications, architecture decisions, domain constraints, tests, process definitions, and design rationale records externalize knowledge and choices for agent work.They provide adequacy criteria, preserve rationale and trajectory, record organization-specific knowledge, define review and autonomy boundaries, and maintain coherence across decisions.Source: §5 · p. 4
- Deterministic process automation was limited to narrow technical phases, whereas AI can navigate incomplete specifications and surface inconsistencies for human judgment.The paper frames AI-enabled process delivery as a new direction for the intent behind Process-centered Software Engineering Environments.Source: §5 · pp. 4–5
6 Conclusions
The paper concludes that software engineering and software engineers are necessary to structure trustworthy AI-agent work. It calls for reframing the replacement debate, investing in engineering practices, and valuing SE education and expertise.
6 Conclusions
- SE defines the structured space for trusted agents, while engineers carry that knowledge into practice by defining agent guidance, eliciting domain meaning, and owning system trajectory.The paper concludes that neither the discipline nor the profession is in decline.Source: §6 · p. 5
- The paper reframes the central question as what structured knowledge must govern agents to produce correct, trustworthy, lasting software, requiring SE and engineers.It argues against treating their abandonment as the answer to whether AI replaces software engineers.Source: §6 · p. 5
- Specifications, domain elicitation, architectural governance, and process discipline are investments in the persistent, coherent space where agents can operate.Organizations treating these practices as optional will pay costs in correctness, maintainability, and accountability.Source: §6 · p. 5
- SE education, research, and practice are prerequisites for AI exploitation and governance, not legacies displaced by AI fluency.The paper warns that prompt-only training without methodological knowledge, domain elicitation, or architectural judgment will reproduce described failure modes at scale.Source: §6 · p. 5