Recoleta
GitHub
Source note

Human-AI Collaboration and the Transformation of Software Engineering Work

Authors
Mamdouh Alenezi
Source
arXiv
Published
Collection
Software Intelligence
Human AI CollaborationAgentic Software EngineeringCode IntelligenceSoftware Engineering RolesAI Governance
Open arXivMarkdown source

Summary

The paper argues that GenAI and coding agents are moving software engineering work toward intent specification, agent direction, verification, and governance. Its main output is a synthesis-based competency model, not a new coding model or benchmark.

Problem

  • AI coding assistants and agents can write, test, review, and submit software changes, so teams need to know what human engineers should still own.
  • The problem matters because faster code production can raise review load, security risk, technical debt, and accountability gaps.
  • Existing evidence is split across productivity studies, repository studies, education papers, and governance work, so the paper tries to connect those findings into one account of changing engineering work.

Approach

  • The authors use a structured interpretive synthesis of recent peer-reviewed and archival sources about GenAI, agentic AI, software engineering roles, and competency needs.
  • They compare three coexisting modes of work: Traditional Software Engineering, Generative AI-Enabled Software Engineering, and Agentic AI-Enabled Software Engineering.
  • They classify engineering activities as automated, augmented, or newly emerging, then map those changes to required human skills.
  • They propose a competency model with five categories: technical, cognitive, socio-technical, governance, and organizational capabilities.
  • They derive nine testable propositions for future empirical work.

Results

  • The paper cites the AIDev study as key evidence: 456,535 agent-authored pull requests across 61,453 repositories and 47,303 developers, produced by five agents: OpenAI Codex, GitHub Copilot, Devin, Cursor, and Claude Code.
  • It reports one AIDev example where a single developer produced 164 agent-authored pull requests in three days, compared with 176 human-authored pull requests over the prior three years.
  • It states that some agents close pull requests up to an order of magnitude faster than humans, while agent-authored pull requests are merged less often and tend to make fewer structural code changes.
  • It cites Peng et al. reporting about 56% faster completion for developers using an AI pair programmer on a standardized task.
  • It cites a workshop study with 22 professional software engineers using ChatGPT for three hours and a survey of 410 developers on AI coding support.
  • Its main claimed contribution is conceptual: three work paradigms, a ten-dimension comparison, a five-category competency model, and nine empirical propositions.

Link