Skip to content

Will Engineers Stop Growing in the AI Agent Era?

Intended audience:

Software development team managers, engineers who use AI coding tools daily, and technology leaders reconsidering their hiring and talent development strategies

Key Takeaways

  • The proliferation of AI coding agents is driving a dual crisis: a sharp decline in junior hiring (new grad hiring ratio down to 7%) and senior skill atrophy (17% drop in comprehension scores)
  • In February 2026, Microsoft's Azure CTO co-authored a paper warning that "chasing short-term efficiency alone will eliminate the next generation of technical leaders" — awareness of this as an industry-wide structural problem is growing
  • The core issue is not "how fast code gets written" but that the pathways for developing judgment are being severed — the response lies in redesigning mentorship structures and intentional skill maintenance

The Big Picture: The Dual Structure of "Not Training" and "Not Growing"

As AI coding agents penetrate software development workflows, two distinct layers of problems are advancing simultaneously.

The first is talent pipeline collapse. Junior engineer hiring is plummeting, narrowing the very pathway through which future seniors emerge. The second is skill atrophy among existing engineers. Research shows that even engineers already performing at senior levels face the risk of core capability degradation through AI dependency.

These two problems are not independent. Fewer juniors means fewer opportunities for seniors to articulate and reaffirm their own knowledge through mentoring. The result is a structural risk where both the entry and exit points of the pipeline deteriorate at the same time.

The Sharp Decline in Junior Hiring: What the Numbers Show

Signs of pipeline collapse have been consistently confirmed across multiple quantitative studies.

According to Stanford University's Digital Economy Study, employment of software developers aged 22–25 declined by approximately 20% from its 2022 peak through September 20251. This decline coincided with the proliferation of AI coding tools, and a correlation was observed showing that occupations with higher AI exposure saw greater employment reductions among younger workers.

A large-scale study by Harvard University researchers tracking 62 million workers and 285,000 companies found that junior-level employment at companies adopting generative AI declined by approximately 7.7% within six quarters of adoption, while senior-level employment remained essentially unchanged2. This asymmetry sits at the heart of the pipeline problem.

Hiring market data points in the same direction. SignalFire's 2025 State of Talent Report showed that the new graduate hiring ratio at major tech companies had fallen to just 7% of total hires in 2024, a 25% year-over-year decline3. Salesforce indicated it might not hire new engineers in 2025, and both Google and Meta have significantly reduced their new graduate hiring ratios since 2022.

The Cascade Effect of "Pulling Up the Ladder"

Cuts to junior hiring inflict serious damage on organizations with a 5–10 year time lag. The industry increasingly refers to this as the Seniority Cliff.

The mechanism is simple but highly irreversible. Companies that did not hire juniors in 2023–2024 will find themselves without mid-level talent in 2025–2026. Existing mid-level engineers, now scarce, command premium prices and get poached externally. The remaining seniors burn out under excessive workloads or leave. With no internal promotion pipeline, companies must rely on external hiring — but the external market itself is shrinking due to the same structural problem.

The insidious aspect of this problem is that it initially appears as a successful efficiency gain. "One senior plus AI agents producing the output of three people" looks excellent on quarterly productivity metrics. The costs only materialize years later, creating a structural incentive problem that leads to flawed management decisions.

Senior Skill Atrophy: The Inconvenient RCT Data

Beyond the pipeline problem, another serious finding has emerged: working under AI assistance may not just fail to maintain existing skills — it may actively degrade them.

Anthropic RCT: 17% Drop in Comprehension

In a randomized controlled trial (RCT) conducted by Anthropic, developers who used AI coding assistance while learning a new coding library scored 17% lower on comprehension tests compared to those who coded manually4. While this data points more to the inhibition of skill formation than "atrophy" per se, the more notable finding was the difference based on usage patterns. Developers who used AI for conceptual questions maintained scores above 65%, while those who delegated code generation to AI scored below 40%. In other words, how AI is used decisively determines learning outcomes.

METR RCT: Even Experienced Developers Were 19% Slower

In a separate RCT conducted by the nonprofit research organization METR, the productivity effects of AI tools were measured on experienced open-source developers working on their own repositories. Contrary to expectations, task completion was 19% slower when using AI tools5. Since developers themselves estimated they were 20% faster with AI, the gap between subjective perception and objective measurement was also striking. METR suggested that when experienced developers are already deeply familiar with their own codebases, the cost of sharing context with AI may exceed the tool's benefits.

Microsoft/CMU Study: Reduced Cognitive Effort for Critical Thinking

A 2025 joint study by Microsoft Research and Carnegie Mellon University found, through self-reported data, that people with higher AI dependency tended to invest less cognitive effort in critical thinking6. The study suggests a paradoxical structure: the thinking habits most needed precisely when AI output requires verification are weakened by everyday AI use.

Skill Atrophy Recognized Within Anthropic Itself

Engineers at AI development companies themselves are beginning to feel this problem firsthand.

According to an internal report published by Anthropic in December 2025, 27% of Claude-assisted tasks were "work that would not have been performed without AI"7. While AI was expanding the scope of work, concerns about skill atrophy had spread among engineers. One engineer reported that over 70% of their work had shifted to code review and corrections, while another imagined a future where their role would involve "being responsible for 1, 5, or perhaps 100 Claudes."

The report noted that workflow changes were also affecting team dynamics. As AI becomes the "first colleague to consult," the incidental learning opportunities that arise from asking teammates questions and conducting code reviews — the spontaneous occasions for mentoring — are diminishing.

The Microsoft Executive Warning: A February 2026 Paper

The clearest expression of industry alarm over this issue came in a paper published in Communications of the ACM (CACM) in February 2026, authored by two Microsoft executives.

The paper "Redefining the Software Engineering Profession for AI," co-authored by Azure CTO Mark Russinovich and Developer Community VP Scott Hanselman, centers on the asymmetry where AI coding agents provide an AI boost to senior engineers but create an AI drag on Early-in-Career (EiC) engineers8. Russinovich described this premise as "a hot topic in every customer engagement, with every company saying they see the same phenomenon internally"9.

The paper's specific failure case is illustrative. An AI agent "solved" a race condition by inserting a delay, masking the root cause of a synchronization bug — a fix that looks plausible to a developer without concurrency experience8. Without experience-backed judgment, engineers cannot properly reject AI output, exposing a fundamental structural problem.

Two countermeasures were proposed. First, a preceptor model — formally pairing seniors and juniors for mentorship that includes AI system operations. Second, implementing an EiC mode in coding assistants that coaches rather than provides answers8. However, the paper itself acknowledges that given AI output quality, the latter cannot always serve as a reliable mentor.

The Case for Optimism — and Its Limits

The data is not all pessimistic. The most influential counterargument comes from AWS CEO Matt Garman. In August 2025, Garman stated flatly that "replacing juniors with AI is one of the dumbest ideas I've ever heard," arguing that juniors are low-cost, adapt quickly to AI tools, and are essential for long-term organizational growth10.

Some data also supports the optimistic view. Stack Overflow's 2025 Developer Survey found that 55.5% of early-career developers used AI tools in their daily work — a higher adoption rate than senior engineers11. The possibility of "reverse mentoring," where juniors as an AI-native generation teach seniors how to leverage AI, has been mentioned across multiple sources.

However, this optimism has structural limits. A high AI tool adoption rate and the formation of deep technical judgment through that adoption are separate problems. Anthropic's finding that "delegating code generation to AI drops comprehension below 40%"4 suggests that merely being able to use AI does not guarantee skill formation.

The Emerging Industry Consensus

Across both optimistic and pessimistic perspectives, the following recognitions are broadly shared within the industry.

AI does not eliminate the junior role but transforms it. The expected skill set is shifting from code generation to AI output verification and design judgment. Stopping junior hiring creates a "Seniority Cliff" 5–10 years later. Virtually no prominent voice disputes this causal relationship.

Even current seniors need intentional "AI-free" practice to maintain their skills, following the same logic as athletes who never skip fundamental drills12. The content of mentoring is also evolving — moving from teaching loop syntax and programming constructs to teaching AI output review methods, architecture decisions, and prompt design.

Most importantly, this problem has the structure of a fallacy of composition: individually rational decisions by each company (reducing juniors and optimizing with AI) produce a collective talent shortage when they occur simultaneously across the industry. Because companies that poach trained talent always exist without bearing training costs, there is a structurally weak incentive for training investment. Economists call this a Poaching Externality13.

Actions Organizations Should Take Now

Based on research data and industry discussion, the measures technology organizations should consider can be summarized as follows.

On the hiring front, maintaining junior hiring slots is the top priority. The long-term value of not severing the pipeline — even at the cost of short-term productivity decline — is consistently emphasized by multiple studies and industry leaders. The preceptor model proposed in Microsoft's paper (formal senior-junior pairing) offers a concrete method for elevating mentoring from individual goodwill to organizational infrastructure8.

For skill maintenance, the key is applying discipline to how AI tools are used. As Anthropic's research demonstrates, using AI as a "conceptual question tool" preserves learning effects, while using it as a "code generation delegate" significantly reduces comprehension4. Setting AI usage guidelines within teams, particularly restricting wholesale code generation delegation for juniors, is a rational approach.

As an organizational design consideration, distributing code review workload is an increasingly critical challenge. One data point shows that AI assistance increased the volume of code produced by juniors, resulting in a 91% increase in senior review time and an 18% expansion in PR size14. Unless review systems are redesigned for AI-era production volumes, senior burnout will accelerate.

Conclusion

The "engineers stop growing" problem brought by AI coding agents has a different structure from simple automation-driven job displacement. Through the dual pathways of juniors losing opportunities to build experience and seniors experiencing skill atrophy, the industry's technical judgment is quietly deteriorating.

The most troublesome characteristic of this problem is that it initially appears as a successful efficiency gain. Whether organizations can make decisions that account for the senior shortage five years ahead — even while quarterly productivity metrics improve — will determine the resilience of technical organizations.

The essential question in software engineering — as Russinovich and Hanselman's formulation makes clear — is not "how much code machines can produce" but "how effectively humans can learn to reason alongside machines"8.

  1. Erik Brynjolfsson, Bharat Chandar, John Chen, "Canaries in the Coal Mine? Six Facts about the Recent Labor Market Effects of AI," Stanford Digital Economy Lab, 2025. Tracked employment trends of workers aged 22–25 using ADP payroll data. Initial version August 2025 (July data), later updated with September 2025 data. https://digitaleconomy.stanford.edu/wp-content/uploads/2025/08/Canaries_BrynjolfssonChandarChen.pdf ↩

  2. Seyed M. Hosseini & Guy Lichtinger (Harvard University), "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data," 2025. Study of 62 million workers and 285,000 companies. DiD analysis reporting 7.7% decline in junior employment (p.4). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555 ↩

  3. SignalFire, "2025 State of Talent Report," 2025. VC report analyzing trends in tech company new graduate hiring ratios. Referenced via Computing.co.uk (February 2026). ↩

  4. Anthropic, "AI Coding Assistance and Skill Formation," 2025. Randomized controlled trial measuring learning effects under AI coding assistance. Referenced via InfoQ (February 2026). https://www.infoq.com/news/2026/02/ai-coding-skill-formation/ ↩↩↩

  5. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," 2025. RCT targeting experienced OSS developers. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ ↩

  6. Microsoft Research & Carnegie Mellon University, "The Impact of Generative AI on Critical Thinking," 2025. Self-reported survey examining the relationship between AI dependency and cognitive effort for critical thinking. Referenced via Addy Osmani, "Avoiding Skill Atrophy in the Age of AI" (April 2025). ↩

  7. Anthropic, Internal AI Usage Report, December 2025. Reported by Interview Query (December 2025). https://www.interviewquery.com/p/anthropic-ai-skill-erosion-report ↩

  8. Mark Russinovich & Scott Hanselman, "Redefining the Software Engineering Profession for AI," Communications of the ACM (CACM), February 2026. https://devops.com/microsoft-executives-warn-ai-could-limit-the-developer-talent-pipeline/ ↩↩↩↩↩

  9. The Register, "Microsoft execs worry AI will eat entry level coding jobs," February 23, 2026. https://www.theregister.com/2026/02/23/microsoft_ai_entry_level_russinovich_hanselman/ ↩

  10. Matt Garman (AWS CEO), public statement, August 2025. Referenced via CodeConductor, "Junior Developers in the Age of AI" (January 2026). ↩

  11. Stack Overflow, "2025 Developer Survey," 2025. Annual survey of over 49,000 developers. https://stackoverflow.blog/2025/12/26/ai-vs-gen-z/ ↩

  12. Addy Osmani, "Avoiding Skill Atrophy in the Age of AI," April 2025. https://addyo.substack.com/p/avoiding-skill-atrophy-in-the-age ↩

  13. IT Pro, "Are we facing an AI-fueled talent pipeline time bomb?", February 2026. Includes reference to economist Teeselink's Poaching Externality concept. https://www.itpro.com/business/careers-and-training/are-we-facing-an-ai-fueled-talent-pipeline-time-bomb ↩

  14. Crossbridge Global Partners, "Senior vs. Junior Developers in the AI Era," January 2026. https://gocrossbridge.com/blog/senior-vs-junior-developers/ ↩