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

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 mixedmethods 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.

Introduction. “...the notion of the programmer as an easily replaceable component in the program production activity has to be abandoned. Instead the programmer must be regarded as a responsible developer and manager of the activity in which the computer is a part.” [64] - Peter Naur, 1985 Implement the feature, add tests, open the PR used to be an engineer’s checklist. Today, Artificial Intelligence (AI) agents can carry out all three, sometimes better, sometimes worse, almost always faster, until it’s not [72]. Industry momentum has been brisk: “vibe coding” has gone mainstream; repositories ship agent metadata (e.g., AGENTS.md) to steer AI toolchains [65]; model providers advertise “learning modes” that retain context across sessions [66]; and some firms are experimenting with AI-enabled interviews that explicitly permit tool use [48]. Concerns have shifted over time: from human error in code, to AI hallucinations, and now to delegation, responsibility, and control when capable agents are part of the workflow. The distinction between generative AI (which produces new content through prompt-response interaction) and agentic AI (which autonomously executes actions) becomes crucial as developers increasingly adopt tools like Cursor and GitHub Copilot that can independently modify and interpret codebases [54]. Recent HCI and SE research suggests that AI will become increasingly integrated into software development, though the partnership remains symbiotic: AI serves as a tool, while human judgment, ingenuity, and creativity are essential to evaluate and complement AI output [4, 58, 83]. Thus, it is critical to understand how agency allocation—the distribution of decision authority and accountability—with AI is distributed across experience levels. In industry, junior engineers enter the workforce alongside AI, navigating the tension between productivity and learning, while senior engineers must adapt mid-career, revising their leadership and mentorship practices. Investigating this generational divide is particularly important because traditional mechanisms for transferring and building expertise—such as pair programming, discussion, and documentation [62, 76]—now need to integrate AI as both a tool and an intermediary. Software engineering has always depended on a talent pipeline where juniors gradually mature into seniors by developing best practices, judgment under uncertainty, and mentoring capacities. AI unsettles this process, especially as industry narratives diverge: some firms are cutting junior roles as “replaceable by AI,” while others prize “AI-native” hires [11, 39, 40, 79]. Prior work has highlighted the importance of investigating both the technical and soft skills required for AI-augmented software engineers [4], as well as the effects on team dynamics [69]. Building on this, it is critical to understand how professional growth, learning, and mentoring align with idealized human-AI collaboration—particularly when companies take extreme approaches to AI adoption. In this study, we zoom in on AI-native junior engineers and senior engineers with pre-AI instincts. We conducted a three-phase qualitative study with junior and senior software engineers to investigate how generative and agentic AI tools shape software practice, professional growth, and mentorship. Our approach combined semistructured interviews, structured elicitation techniques (Applied Cognitive Task Analysis (ACTA) [61], Delphi method [29, 34]), and task-based activities (AI-assisted debugging and prompt review). In Phase 1, we interviewed five senior engineers with ACTA to elicit examples of tacit knowledge [16, 82], refining these through Delphi-inspired consensus-building to converge on a final, realistic debugging task. In Phase 2, 10 junior engineers engaged in the Phase 1 final task using Cursor, an agentic AI tool, followed by postmortems, surveys, and reflective interviews on their AI use. In Phase 3, five different senior engineers reviewed anonymized junior artifacts—including code, prompt histories, and reflections— mirroring realistic mentorship contexts and evaluating how AI records support knowledge transfer. Across these phases, we also conducted semi-structured interviews, enabling us to triangulate how engineers allocate agency with AI, perceive career development, and approach mentorship or learning in AI-mediated workflows. Our findings highlight that agency allocation in AI-mediated software work is preconfigured at the organizational layer (policies, tooling defaults, repos, CI guardrails) before individual preferences matter. Within those bounds, seniors and juniors take varying routes to preserve control with the way they engage with generative and agentic AI.

Related work. 2.1 Generative/Agentic AI in the Software Industry Large Language Models (LLMs), a type of Generative AI, produce new content typically accessed through a prompt-and-response interaction. Agentic AI, however, goes a step further: rather than leaving the programmer to act on an output, it can autonomously execute actions in response to a prompt, with the programmer primarily reviewing, approving, or rejecting changes. Put differently, agents are autonomous, iterative systems that perceive feedback and operate in dynamic environments, whereas a standalone LLM is not, by itself, an agent. Consequently, software engineers are increasingly integrating agentic AI into their workflows when compatible [15, 50]. Popular examples—such as GitHub Copilot, Cursor AI, and Windsurf [1–3]—seamlessly embed into developer IDEs [54]. Agentic AI also supports “vibe coding,” where high-level natural language prompts lead to the evaluation, generation, and deployment of code [71], shifting greater control to the agentic system and reducing the user’s control [77]. Prior research has examined how both software engineers and students engage with generative and agentic AI across a range of contexts. Studies of expert developers have explored how they use AI to perform software tasks [49, 88, 90], while investigations of junior engineers highlight challenges such as reduced learning opportunities alongside productivity gains [80]. Studies have also surfaced students’ and professionals’ interactions with coding agents across experience levels [7, 15], and suggested design guidelines and collaboration frameworks to enhance AI use [30, 55]. Building on this foundation, recent work recommends further exploratory studies to understand how groups with different levels of expertise experience and take part in AI adoption [24], especially as agentic AI tools evolve and become increasingly widespread. Thus, we extend prior work by investigating how junior and senior engineers engage with both generative and agentic AI in their day to day jobs. Our study looks at two ends of the spectrum: engineers who began their careers alongside AI and those who spent most of their careers without it. Rather than detailing specific workflows, we focus on how these groups preserve control, balance productivity and learning, and experience psychosocial impacts such as imposter syndrome.

2.2 Agency for Software Engineers Broadly, agency is defined as the ability to act driven primarily by internal thoughts and feelings, rather than the external environment [20, 84]. In a professional sense, Eteläpelto et al. defines professional agency as exercised when “professional subjects and/or communities influence, make choices, and take stances on their work and professional identities” [23].

Method. Overview We designed a three-phase study with both senior and junior software engineers, combining semi-structured interviews, structured elicitation techniques (Delphi Method, ACTA), and task based activities (Cursor debugging task and prompt reviews). With these methods, we aimed to answer the following research questions:

(1) RQ1: How do junior and senior software engineers allocate agency 1 between themselves and agentic / generative AI in daily work? (2) RQ2: How do junior and senior software engineers perceive professional growth for a junior in the age of AI? (3) RQ3: When and why do engineers deem mentorship indispensable in workflows with agentic / generative AI? (4) RQ4: How do AI records (e.g., prompt history, provenance) shape code review and mentorship?

3.1 Participants We recruited 20 professional software engineers (10 juniors, 10 seniors) across three phases using convenience and snowball sampling (LinkedIn outreach and participant referrals). We defined seniors as engineers with ≥5 years full-time experience and ≥1 year in an advanced role (senior, staff, or lead), and juniors as engineers with ≤1 year full-time experience. Recruitment targeted a spread of company types (e.g., Big Tech/Cloud, FinTech/Financial Services, Enterprise SaaS, DevTools, HealthTech) to capture variability in AI policies and tooling. Our goal was qualitative, exploratory insight across 1-hour long semi-structured and task based interviews into how agency and mentorship shift with agentic AI across experience levels. We therefore prioritized depth over breadth in a three-phase design detailed below.

3.2 All Phases 3.2.1 Pre-Interview Survey. Prior to interviews, all participants completed a survey capturing demographic information, years of professional experience, and details regarding the types and usage of AI-assisted tools.

3.3 Phase 1 (P1) We conducted 60-minute interviews with 5 senior software engineers (all male, aged 28–46, based in the US), supplemented by one survey, all administered remotely via Zoom.

3.3.1 Semi-Structured Interview - 20 minutes. We introduced a working definition of tacit knowledge and asked engineers to reflect on how they had acquired tacit expertise, how AI-assisted coding tools and IDEs (e.g., GitHub Copilot, ChatGPT, Cursor) affect their workflows, and their mentorship practices with AI.

3.3.2 Task-Based Elicitation - 40 minutes. To capture concrete examples of tacit expertise, we integrated task-based elicitation into the interviews. Using the ACTA framework (Task diagram, Knowledge Audit, Simulation), participants decomposed domain-relevant scenarios (e.g., debugging or feature development) into steps. For each step, they identified cognitively demanding judgments and reflected on the risks and benefits of delegating these steps to AI, some of which are shown in Appendix A.1. This approach surfaced “hidden” reasoning strategies often overlooked in interviews alone. We employed ACTA [61] to systematically elicit the implicit (i.e. tacit) knowledge and cognitive strategies that distinguish senior from junior engineers—expertise developed over years that shapes how seniors approach complex problems differently. While traditional structured interviews might capture what engineers do, ACTA’s structured task decomposition reveals how they think through problems, identify critical decision points, and recognize patterns that novices often miss. This method was particularly valuable for generating authentic debugging scenarios for Phase 2, as it surfaces the “hidden” cognitive demands within routine tasks that seniors navigate intuitively but juniors may struggle with, especially when relying on AI assistance. By having seniors explicitly articulate which aspects of tasks require judgment versus which could be delegated to AI, ACTA helped us identify tasks that would reveal meaningful differences in how juniors and seniors allocate agency (RQ1) and approach problem-solving with AI tools.

Discussion. Here, we synthesize P1 and P3 seniors’ boundary descriptions and P2 juniors’ on-task decisions and descriptions of their daily practices. Our results give clear examples of who initiates changes when working with generative and agentic AI, when control returns to the human, and how engineers ensure they are answerable for their code.

4.1.1 Company rules around AI.

Company policies preconfigure AI agency boundaries before individual preferences matter. Across phases, participants described mandates, allow-lists, and security constraints that preconfigure who writes what-an agent or a person-before individual tool preference ever enters. Companies varied in their rules for AI tool usage. Some had formal restrictions, such as limiting engineers to an approved internal set of tools with contracts in place (J8, J6, J4, S7, S3, S1, S10), prohibiting the use of non-company AI on corporate devices (S8), or requiring that confidential data not be shared with external services (J3, J10, J4, S6, S2, S1). In tightly regulated settings, participants named data to protect from AI-“Social Security numbers, credit card numbers, personally identifiable information” (S2), echoing that their companies do not allow agentic AI and use old versions of Claude and ChatGPT (S2, S6, S10). Overall, AI tools were concrete and varied, as show in Table 1, 2, 3. Companies also deployed internal knowledge bots, assistants that turned specs into code, Notebook LLMs, AI diagramming aids, and public prototyping tools like AI Studio (S8, J6, S7). In some cases, employees were required to use specific tools (e.g., Cursor) on a regular basis (J5, J9, S5), while in others there were no clear or memorable policies (J1, J2, S9, S4). Most companies actively pushed engineers to use AI (Figure 3a), while there was mixed knowledge on if their company has access to their prompt history (Figure 3b). Several (S8, S5, S4, S1, J9, J7) described a subtle loss of agency in the “use AI now” push from top-down, delivered matter-of-factly rather than as a choice. Within companies, seniors and juniors mentioned various capabilities that were tacit to their role that limited their use of AI, as often AI can not reliably infer infrastructure conventions, tooling, and history. S3 offered an example: AI that proposes a virtualenv-based training pipeline that “runs” but cannot be checked in due to incompatibilities with company specific Bazel and Pytorch versions. S1 added that a product’s “true nature” often lives in distributed mental models; their team’s mitigation is to externalize key knowledge so progress is not person-dependent. Similarly, participants (J9, J8, J7, J6, S9, S2, S10, S8, S4) raised concerns about answerability for AI-generated code, with J9 noting, “there’s a lot of code ... [seniors] have to trust that you tested ... sometimes you wake up and something is broken”. Overall, seniors reported high self-confidence in coding without AI, more varied views on AI’s usefulness, and strong confidence in evaluating AI output (Figure 3d, Figure 3c, Figure 3e). After completing the coding task with agentic AI, juniors demonstrated varied responses in their self-assessed confidence for “Confidence in ability to code without AI” (Pre-Task M= 3.90 ± 0.88) to Post-Task M= 3.30 ± 0.82), Cohen’s d= 0.71). The majority (6 of 10 participants) reported decreased confidence post-task, with changes ranging from −1 to −2 points on the 5-point scale. This trend was particularly notable among those who initially rated themselves highest (at level 5): all three of these participants decreased their ratings post-task. In contrast, self-confidence in evaluating AI output had diverse trajectories pre and post task: 3 participants decreased their ratings, 4 increased, and 3 remained stable.

Conclusion. Our study examines how agentic and generative AI integrate into software engineering, shaping agency, expertise, and mentorship across experience levels. While AI promises productivity gains, its impact on professional development is uneven: senior engineers leverage foundational instincts to maintain strategic control, whereas junior engineers gain exposure but risk fragile understanding and impostor syndrome. Crucially, AI cannot replace the tacit knowledge transfer essential for system thinking, failure anticipation, and architectural judgment. To address these challenges, we highlight three evolving practices. First, Preserving Individual Agency: company-wide and personal practices for using AI tools— such as incremental changes, interrupting, and verifying outputs— help engineers maintain control and responsibility over their work. Second, Evolving the Mentorship Pipeline: collaborative and individual practices in which senior engineers pass down intuition, critical thinking, and judgment to junior engineers support their learning, agency, and professional growth in an AI-mediated environment. Third, Prompt & Code Reviews (PCRs): a collaborative practice to maintain agency by making AI interactions accountable. Juniors document and justify key prompts, while seniors oversee accountability, ensuring AI-native juniors remain authors of their reasoning and retain ownership over outputs. Overall, preserving agency in software development, learning, and mentorship is essential to maintain a sustainable talent pipeline and ensure engineers can guide, not just operate, increasingly autonomous systems.

Limitations. 5.4 Limitations & Future Directions While our study provides exploratory insights into how engineers of different experience levels interact with AI in software development, several limitations warrant careful consideration. Our findings represent a snapshot in time (the summer of 2025) to help researchers understand software engineers at the cusp of widespread adoption of AI-assisted tools.