From Junior to Senior: Allocating Agency and Navigating Professional Growth in Agentic AI-Mediated Software Engineering

TL;DR

The paper investigates how junior and senior software engineers allocate agency when interacting with agentic and generative AI in real-world workflows. Using a three-phase mixed-methods design (ACTA+Delphi with seniors, AI-assisted debugging with juniors, and prompt-history reviews by seniors), it reveals that organizational policies preconfigure AI-usage boundaries and that seniors maintain control through targeted delegation while juniors struggle with ownership and overreliance. It contributions three practices—Preserving Individual Agency, Evolving the Mentorship Pipeline, and Prompt & Code Reviews (PCRs)—to preserve human agency, lesson AI-induced deskilling, and sustain a robust talent pipeline. The findings have practical implications for governance, mentorship, and tool design in AI-mediated software development, supporting safer, more effective human–AI collaboration at scale.

Abstract

Juniors enter as AI-natives, seniors adapted mid-career. AI is not just changing how engineers code-it is reshaping who holds agency across work and professional growth. We contribute junior-senior accounts on their usage of agentic AI through a three-phase mixed-methods study: ACTA combined with a Delphi process with 5 seniors, an AI-assisted debugging task with 10 juniors, and blind reviews of junior prompt histories by 5 more seniors. We found that agency in software engineering is primarily constrained by organizational policies rather than individual preferences, with experienced developers maintaining control through detailed delegation while novices struggle between over-reliance and cautious avoidance. Seniors leverage pre-AI foundational instincts to steer modern tools and possess valuable perspectives for mentoring juniors in their early AI-encouraged career development. From synthesis of results, we suggest three practices that focus on preserving agency in software engineering for coding, learning, and mentorship, especially as AI grows increasingly autonomous.

Figures (9)

• Figure 1: Debugging task interface showing the React-based admin panel application with multiple bugs across different pages. The User Activity Logs page (a) displays user authentication and activity tracking with pagination controls. The Analytics Dashboard (b) presents key metrics (users, revenue, sessions, conversion rate), trend analysis, top page performance, and user behavior statistics.
• Figure 2: Survey responses across all three phases regarding AI adoption and confidence. Comparative bar charts showing responses from Phase 1 seniors (n=5), Phase 2 juniors (n=10), and Phase 3 seniors (n=5). (a) Company push: 16 of 20 participants reported their companies actively incentivize AI tool usage. (b) Access to prompts: Mixed awareness about company access to prompt history, with plurality responding "Don't Know" (9/20). (c) AI helpfulness: Confidence in generative AI's ability to help with coding tasks, rated 1-5, showing moderate confidence across groups. (d) Confidence in self: Self-rated ability to complete coding tasks without AI, with seniors showing notably higher confidence (median 5) than juniors (median 3-4). (e) AI evaluation: Confidence in evaluating AI-generated code, where seniors unanimously rated 3-5 while juniors showed more variation (2-5).
• Figure 3: Results from Phase 2 (P2). Figure showing debugging task data for 10 junior engineers. Columns show participant ID from J1 to J10, detailed prompt sequences listing each interaction mode and type, task duration in minutes, and completion status. Prompt types include debugging, code explanation, codebase understanding, React concepts, and implementation suggestions. Task durations varied from 7 to 30 minutes with different completion levels.
• Figure 4: Post-task survey responses from Phase 2 junior engineers. Five bar charts displaying participants' self-reported assessments immediately after completing the 30-minute debugging task. Confidence in fix (1-5 scale): Most rated 3-4, indicating moderate confidence in their solutions despite varying completion rates. Current codebase understanding (1-5 scale): Half selected 2, with none rating 5, reflecting limited comprehension of the unfamiliar React application after the brief task. Mental investment (1-5 scale): Predominantly 4 ratings indicate high cognitive engagement despite AI assistance. Importance to do well (1-5 scale): Most rated 4-5, demonstrating strong motivation to succeed in the research task. Hard work required (1-3 scale): Most rated 2-3, suggesting the task demanded significant effort. These results highlight that while juniors remained engaged and motivated, they struggled with confidence and understanding when debugging unfamiliar codebases, even with agentic AI support.
• Figure 5: Senior engineer S1's Applied Cognitive Task Analysis (ACTA) decomposition of debugging workflows. Task diagram showing S1's breakdown of debugging steps with annotations for cognitive demands and AI delegation considerations. S1's analysis illustrates how experienced engineers conceptualize the balance between maintaining control over critical decisions and leveraging AI for routine subtasks.