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

Agency in AI-mediated software engineering is largely determined by organizational policies, tooling defaults, and repository-level guardrails rather than by individual engineer preferences. Within those constraints, senior and junior engineers allocate agency differently: seniors preserve control via detailed, strategic delegation and tacit judgment, while juniors—who are AI-native—oscillate between over-reliance on agentic AI and cautious avoidance. This dynamic threatens the traditional pipeline for tacit knowledge transfer, making senior mentorship and new institutional practices (e.g., Prompt & Code Reviews) critical to preserve learning, accountability, and professional development.

Key Points

• Study design: three-phase mixed-methods study (qualitative emphasis) with 20 professional engineers:
  • Phase 1: ACTA + Delphi-inspired elicitation with 5 senior engineers to surface tacit knowledge and produce a realistic debugging task.
  • Phase 2: AI-assisted debugging task using an agentic tool (Cursor) with 10 junior engineers; postmortems, surveys, and interviews.
  • Phase 3: Blind review of juniors’ prompt histories, code, and reflections by 5 different seniors to simulate mentorship/code-review contexts.
• Tools and context: Focused on agentic and generative models embedded in IDEs (examples: Cursor, GitHub Copilot); task centered on realistic React maintenance/debugging.
• Core behavioral patterns:
  • High-familiarity tasks: both juniors and seniors use small, task-focused prompts and maintain control through accept/reject decisions.
  • Low-familiarity tasks: seniors use AI strategically (idea generation, scaffolding, low-level work) while retaining judgment; juniors either over-delegate to agents (risking unrecognized hallucinations) or avoid AI to escape uncertainty and responsibility.
• Tacit knowledge and learning:
  • Seniors draw on pre-AI mental models and diagnostic heuristics to steer agents and to critique outputs.
  • Juniors report productivity gains but also impostor syndrome, shallow understanding of changes, and lower sense of accomplishment when relying on agents.
• Accountability and records:
  • AI artifacts (prompt histories, provenance) are useful for mentorship and auditing but require intentional practices to ensure they support learning rather than simply documenting delegation.
• Recommended practices emerging from synthesis:
  • Preserve Individual Agency: incremental, verifiable AI use and interrupt/verify patterns.
  • Evolve Mentorship: seniors intentionally teach judgment, heuristics, and critical evaluation of AI outputs.
  • Prompt & Code Reviews (PCRs): structured collaborative reviews of prompts and agent outputs to maintain accountability and knowledge transfer.
• Limitations: small, convenience sample; primarily U.S.-based participants; qualitative and exploratory—findings suggestive rather than generalizable.

Data & Methods

• Participants: 20 professional software engineers split evenly by seniority (10 juniors with ≤1 year experience; 10 seniors with ≥5 years experience and advanced roles). Companies spanned Big Tech, FinTech, Enterprise SaaS, DevTools, HealthTech, etc.
• Phase 1 (Seniors, n=5):
  • 60-minute sessions combining 20-min semi-structured interviews and 40-min ACTA (Applied Cognitive Task Analysis) elicitation to surface tacit decision points.
  • Followed by a Delphi-inspired consensus survey to select/shape a representative debugging task.
• Phase 2 (Juniors, n=10):
  • 65-minute sessions performing the consensus debugging/maintenance task in a React codebase using Cursor in Agent or assistive modes.
  • Data collected: prompt histories, code changes, session recordings, post-task surveys, and reflective interviews.
• Phase 3 (Seniors as reviewers, n=5):
  • Blind review of anonymized junior artifacts: prompts, diffed code, and postmortems; reviewers assessed how artifacts supported mentorship and quality assurance.
• Analysis: Qualitative coding and triangulation across ACTA outputs, task performance, interviews, and prompt/code review reactions to infer patterns in agency allocation, learning, and mentorship needs.

Implications for AI Economics

• Labor demand and task composition
  • Routine, well-defined tasks are increasingly automatable by agentic AI, raising the marginal productivity of juniors on those tasks but potentially reducing the value of acquiring those routine skills. This can lead to task reallocation toward higher-level judgment work—augmenting demand for senior-level tacit expertise.
  • Agentic AI may drive task polarization: automation of modular coding work combined with sustained or increased demand for oversight, architecture, and mentorship.
• Human capital formation and career progression
  • The traditional apprenticeship pipeline (learning by doing, code review, pair programming) is disrupted: juniors gain output but less tacit understanding. That may slow accumulation of firm-specific human capital unless firms invest in structured mentorship practices (e.g., PCRs).
  • Firms that under-invest in mentorship risk a cohort of workers with high short-term productivity but lower long-run promotability, increasing turnover or requiring later remediation training.
• Wage structure and hiring strategies
  • Short-term productivity gains from AI-native juniors could compress entry-level wage premia or change hiring signals (valuing AI-tool fluency). However, scarcity of tacit judgment skills may raise the wage premium for experienced engineers who can supervise and audit agentic outputs.
  • Firms may differentially value “AI-native” hires vs. experienced seniors; compensation strategies might shift to reward mentorship, review, and governance work more than before.
• Firm-level heterogeneity & organizational policy effects
  • Because organizational policies, tooling defaults, and CI/Repo guardrails preconfigure agency, firm-level adoption choices will create heterogeneity in realized productivity and learning outcomes. Firms that adopt aggressive automation without guardrails may see faster short-run throughput but face higher risk of latent defects and reduced human capital formation.
  • Investments in provenance, prompt logging, and structured review (which have non-trivial administrative costs) may become necessary to manage liability and product quality—creating a new category of governance costs influencing adoption decisions.
• Liability, regulation, and measurement
  • The presence of agentic actions shifts accountability and raises regulatory concerns (audit trails, explainability). Compliance costs and legal risk allocation will factor into firms’ adoption and labor decisions.
  • Economists and firms need better productivity metrics that disentangle output created by AI from human learning—standard output-per-hour measures may overstate sustainable human productivity.
• Policy and managerial implications
  • Firms should evaluate ROI of mentorship and structured prompt/code-review practices; subsidies or tax incentives for training might be warranted to avoid under-provision of tacit-skill transmission.
  • Labor market studies should measure promotion rates, skill accumulation, and wage trajectories for AI-native cohorts versus pre-AI cohorts to assess long-run effects.
  • Standardizing prompt & provenance practices (industry norms, tooling standards) could reduce frictions and externalities from uneven adoption.
• Research opportunities
  • Quantify the trade-off between short-term productivity (via agents) and long-run human capital accumulation.
  • Measure how differing organizational policies alter hiring, wages, and defect rates across firms and industries.
  • Evaluate the economic impact of formalizing mentorship roles (e.g., payment for review/mentoring time, changes to promotion criteria) and the cost-effectiveness of PCRs and provenance systems.

Limitations to consider when applying these implications: the study is qualitative and exploratory with a small, convenience sample; cross-industry generalization requires larger-scale quantitative work. Nonetheless, the observed mechanisms—organizational preconfiguration of agency, divergence in junior vs. senior interactions with agentic AI, and the centrality of mentorship and records—point to important labor-market and firm-organization dynamics that merit economic measurement and policy attention.