AI tools are reshaping who holds agency in software engineering: senior engineers preserve control through detailed delegation and mentorship, while juniors swing between dependence and avoidance; organizational policies, not personal preference, are the main constraint on agency.

Main Finding

AI is reshaping agency in software engineering: organizational policies, not individual preferences, primarily constrain who can exercise control. Senior developers preserve agency by delegating precisely and leveraging pre-AI instincts; junior developers—who enter as “AI‑natives”—tend to oscillate between over-reliance and avoidance. Seniors’ perspectives are crucial for mentoring and for preserving human control as AI becomes more autonomous.

Key Points

• Agency shifts are sociotechnical: organizational rules and norms (access, review, responsibility) matter more than individual tastes.
• Seniors adapt mid-career by learning to specify and steer agentic AI, maintaining authority through detailed delegation and oversight.
• Juniors are comfortable using AI tools but struggle with when to trust them, leading to two failure modes: over-reliance (accepting unsafe recommendations) and cautious avoidance (under-using valuable automation).
• Seniors’ foundational instincts (debugging heuristics, risk judgment, design thinking) remain valuable and complementary to AI capabilities.
• Blind review of junior prompt histories shows seniors can identify errors and mentoring opportunities, indicating concrete channels for knowledge transfer.
• The authors propose three practice areas oriented to preserving agency: (1) coding (designing delegation patterns and checks), (2) learning (structured curricula/feedback on AI use), and (3) mentorship (senior-guided prompt design and review).

Data & Methods

• Mixed-methods, three-phase design:
• ACTA (Applied Cognitive Task Analysis) paired with a Delphi process involving 5 senior developers to surface expert judgments about AI’s impact on tasks and agency.
• An AI-assisted debugging task with 10 junior developers to observe in-situ interactions and failure modes while using agentic tools.
• Blind reviews: 5 additional seniors evaluated anonymized junior prompt histories to assess common errors and mentoring leverage points.
• Triangulation: qualitative interviews/ACTA, controlled task observation, and external expert review to synthesize behavioral patterns and normative recommendations.
• Sample size and domain: small, focused on software engineering practitioners; emphasis on depth and triangulation rather than broad representativeness.

Implications for AI Economics

• Skill-biased augmentation and returns to experience:
  • Seniors’ ability to steer AI suggests complementarity between experience and AI—potentially increasing wage premiums for experienced engineers who can preserve and exercise agency.
  • Juniors’ over-reliance risks deskilling if organizations don’t invest in guided learning, shifting returns from hands-on coding skills toward meta-skills (prompting, oversight).
• Organizational policy as a key margin:
  • Firm-level rules (access, review requirements, accountability structures) will shape who captures productivity gains from AI. Policies that centralize control can concentrate rents; decentralized, mentorship-oriented policies distribute benefits and human capital accumulation.
• Labor supply, training, and mobility:
  • Investment in mentorship and structured AI-use training can accelerate junior human capital accumulation, affecting career ladders, hiring demand for senior mentors, and the pace of automation-induced displacement.
• Productivity measurement and task allocation:
  • Economic models should account for task-level delegation patterns and the value of “steering” skills; output measures must separate raw code generation from quality, maintainability, and safety that seniors protect.
• Policy/regulatory considerations:
  • Encouraging practices that preserve human agency (audit trails, review protocols, required human-in-the-loop for critical code) could mitigate risks of over-automation and unequal gains.
• Research implications:
  • Need for larger-scale, longitudinal studies linking changes in AI-use practices to wages, promotion rates, and firm performance to quantify distributional effects and dynamic returns to experience.

Limitations to keep in mind: small, domain-specific sample; qualitative emphasis limits causal claims about labor-market outcomes—follow-up empirical work is needed to measure macroeconomic impacts.