Recoleta
GitHub
Source note

Rethinking Software Engineering for Agentic AI Systems

Authors
Mamdouh Alenezi
Source
arXiv
Published
Collection
Software Intelligence
Agentic AISoftware EngineeringMulti Agent OrchestrationCode VerificationHuman AI Collaboration
Open arXivMarkdown source

Summary

This paper argues that software engineering should shift from manual code writing toward orchestrating AI agents, verifying AI-generated outputs, and keeping humans accountable for decisions. It is a structured literature review and position paper, not a new algorithm or benchmark study.

Problem

  • LLMs and agentic systems can generate large amounts of working code, so manual code authorship is no longer the main bottleneck in many workflows.
  • That creates a new bottleneck: checking whether AI-generated code is correct, secure, maintainable, and aligned with requirements.
  • If software engineering practice, education, and governance stay centered on writing code, they will mismatch how AI-assisted development actually works.

Approach

  • The paper conducts a structured multivocal literature review of 23 recent sources, including peer-reviewed studies, preprints, and expert surveys.
  • It organizes the evidence into four themes: engineer role changes, verification challenges, empirical effects on productivity and quality, and implications for education and practice.
  • From that synthesis, it proposes a conceptual model of software engineering built around four competencies: intent articulation, systematic verification, multi-agent orchestration, and accountable human judgment.
  • It also outlines changes to curricula, tools, lifecycle processes, and professional governance so teams can operate in a verification-first, human-in-the-loop development model.

Results

  • The main claim is conceptual: software engineering should be reorganized around orchestration, verification, and human-AI collaboration because code generation is becoming abundant while semantic validation remains hard.
  • The paper itself does not report new experimental results or a new benchmark. Its evidence comes from prior studies.
  • Cited evidence includes a controlled experiment with 151 professional developers where AI assistance reduced task completion time by a 30.7% median in the initial phase, but later maintenance showed no significant difference in completion time or code quality versus control.
  • A cited evaluation found that systems with similar success rates could differ by 53× in cost and by tool-call counts of 917 vs. 3, which the authors use to argue that orchestration quality matters.
  • A cited human-AI collaboration study reports that humans or AI alone solved complex tasks at 0.67% to 18.89% success, while collaboration reached 31.11%.
  • Based on this synthesis, the paper claims the durable role of engineers shifts upward toward system design, verification infrastructure, risk ownership, and oversight of multi-agent workflows.

Link