We all fell for it…

TL;DR

AI coding tools are creating cognitive debt by letting developers skip the painful learning process, atrophying the very skills needed to supervise AI-generated code.

Key Points

• 1.Agentic coding is described as a trap due to cognitive debt. Article author Lars Fay argues that AI coding workflows put growing distance between developers and the code being generated, eroding critical thinking and deep system understanding.
• 2.AI coding causes three quantifiable trade-offs. These are: increased system complexity from AI non-determinism, skill atrophy across wide swaths of developers, and vendor lock-in with fluctuating token costs — Cloud Code outages have already halted entire teams.
• 3.The cost-per-intelligence metric is actually dropping rapidly. GPT-5.5 Medium matches GPT-4.5x High intelligence at less than half the cost ($1,200 vs $2,800 benchmark run), and GPT-5.5 Low matches Claude Sonnet 4.6 at one-eighth the price ($500 vs $4,200).
• 4.AI is a slot machine that removes the productive pain of learning. Theo argues that struggling through unfamiliar code — like learning to ollie in skateboarding — builds foundational skills, and AI lets developers pull a lever instead of enduring that discomfort.
• 5.Natural language prompting is NOT just another abstraction layer. Unlike moving from assembly to C++ or JavaScript, developers who adopt agentic coding report brain fog and skill loss — previous abstraction shifts never produced these cognitive side effects.
• 6.Anthropic's own research identified the paradox of supervision. Effectively using Claude requires coding skills to supervise it, but overusing Claude causes those exact coding skills to atrophy — a self-defeating loop documented in their internal study.
• 7.Senior engineers are not immune — Django creator Simon Willison reported losing mental models. With 30 years of experience, he no longer has a firm understanding of what his own applications do, making each new feature harder to reason about.
• 8.The skill gap between great and average developers has massively widened. Bad developers using AI go from mediocre to net-negative on teams; junior developers hired to code with AI can ship fast but cannot debug anything they didn't generate themselves.
• 9.LLMs accelerate the wrong parts of development — speed over understanding. A good developer's priority list should be: understanding, alignment with standards, conciseness, then speed; AI inverts this, giving speed to developers who haven't earned it through competence.
• 10.Planning through code is lost when agents replace hands-on coding. Theo and Dax (SST/OpenCode) both argue that typing code is how developers discover what they should build — plan mode and spec-driven development miss this exploratory thinking-through-doing process entirely.