Is Agentic Coding a Trap? The Truth About AI Programming Causing Skill Atrophy — and How to Fight Back

The AI Programming Craze and Its Hidden Concerns: An Article That Set the Tech World on Fire

Programmer and entrepreneur Lars Faye's article Agentic Coding is a Trap is sparking heated debate across the tech community. Renowned deep work advocate Cal Newport broke down the article in detail on his show, revealing a deeply unsettling paradox lurking behind AI programming tools.

Lars Faye describes the typical workflow of AI-driven software development today: developers define project requirements, generate plans, and then repeatedly "pull the slot machine lever" — iterating with multiple AI agent instances until the task is complete. Throughout this process, the coordinator drifts further and further from the actual generated code.

This new paradigm, known as Agentic Coding, refers to a workflow where AI autonomously plans and executes multi-step programming tasks — a significant departure from earlier code completion or single-turn Q&A-style AI assistance. Tools like Claude Code, Devin, and GitHub Copilot Workspace can independently read and write files, execute terminal commands, run tests, and iteratively revise based on results — all with minimal human intervention at each step. At the core of this paradigm is the "Tool Use Chain": the LLM is no longer just generating text but acting as a decision-making hub that orchestrates multiple tools to accomplish complex tasks. It is precisely this autonomy that dramatically widens the cognitive distance between the developer and the code.

This new paradigm has sent the entire industry into a frenzy. Some developers claim that with Claude Code they've become "10x developers," even feeling like "100x." The experience of having AI complete complex tasks in a week or debug issues in three days is described as "unbelievable."

But is this vision accurate? Lars Faye's answer is no.

AI Programming Is Eroding the Very Skills Developers Need to Use AI

The Skill Atrophy Paradox of Agentic Coding

Faye raises a critical insight: there's a hidden prerequisite for successfully using agentic coding — only seasoned developers with critical thinking skills and the ability to reason at the architectural level can spot problems buried in thousands of lines of AI-generated code. The irony is that prolonged reliance on AI tools has been shown to erode precisely those capabilities: critical thinking and cognitive clarity.

This concern is backed by solid neuroscience. Skill Atrophy follows the brain's fundamental "use it or lose it" principle: neural pathways that go unused weaken through synaptic pruning. Cognitive psychologists call this phenomenon "Cognitive Offloading" — when external tools consistently handle tasks the brain would otherwise perform, the brain reduces resource allocation to those abilities. GPS navigation weakening spatial memory and calculators diminishing mental arithmetic skills are classic examples. The skill atrophy caused by AI programming tools is essentially the same mechanism amplified across far more complex cognitive domains. Because programming involves systematic logical reasoning, the impact may be far more profound than the degradation of any single skill.

This isn't an abstract worry — it's happening at scale right now. Faye cites numerous Reddit posts as evidence:

• "Losing coding ability because of AI": "I feel like my skills are genuinely deteriorating. Not understanding the syntax itself is the problem. Checking whether the AI has created chaos or subtle bugs takes a lot of time, and you need to understand the language to do that."
• "Feel like I'm getting dumber after using AI": "Every time I use AI for anything, I feel like I'm losing skills and becoming a worse professional."
• "Letting AI do 100% of the coding fried my brain": "Recently I've found myself struggling with general cognitive ability and problem-solving."

Junior Developers Face an Even More Severe Skill Degradation Crisis

For junior developers, the problem is especially acute. Faye points out that the struggle of writing code from scratch is precisely the process that builds code comprehension and review skills. If you skip the struggle and jump straight to code review, you have no solid skills to fall back on.

Related Reddit posts are alarming:

• "The junior devs on our team deliver fast, but can't debug any code they didn't personally write"
• "Hired a junior who learned to code with AI — can't debug without it, don't know how to help them"

In computer science education, this phenomenon is known as the "Junior Year Wall" — after students use AI to breeze through foundational programming courses, the consequences of lacking fundamental understanding hit all at once when they face advanced coursework in their junior year. The root cause is that introductory courses can often be solved through pattern matching and memorization, while advanced courses demand truly internalized computational thinking. AI tools cause this wall to appear earlier and make it harder to detect — students can deliver "working code" but lack genuine understanding of code behavior. A 2023 Stanford study found that after introducing AI-assisted tools, the surface quality of student code improved, but performance in oral examinations and whiteboard coding tests declined significantly, confirming the prevalence of this "illusion of understanding." Faye argues that this effect is now playing out at massive scale across the entire programming industry.

Four Deep Observations from a Developer with 30 Years of Experience

Cal Newport consulted a veteran developer with 30 years of experience, who shared four remarkably insightful observations:

First, AI programming tools genuinely work — but with strict prerequisites. If you're a highly skilled developer and you've established good rules and guidance frameworks, these tools are indeed much faster than manual coding while still producing high-quality code. This is why he continues to use them despite seeing many problems.

Second, attention fragmentation is a severely underestimated problem. Agentic systems shouldn't run unsupervised for long periods, but they're also not instantaneous — each operation involves a one-to-two-minute wait. Your choices are either to sit idle and wait, or to context-switch to other tasks, and the latter leads to runaway context switching, lower-quality output, and chronic mental fatigue.

The cognitive cost of Context Switching is well-supported by research. Psychologist Gerald Weinberg's research shows that working on two projects simultaneously reduces effective time on each by about 20%, and this loss climbs to 40% with three projects. More critically, each task switch requires the brain to reload its "working memory context," a process that takes an average of 23 minutes to fully restore deep focus (Gloria Mark, UC Irvine research). The "wait-switch-wait" pattern created by agentic programming tools