Agentic Coding Is a Trap: Cognitive Debt and Atrophy

Lars Faye's article "Agentic Coding is a Trap" challenges the hype around Spec Driven Development (SDD) and AI coding agents. He warns that while tools like Claude Code are powerful, they introduce cognitive debt, skill atrophy, and vendor lock-in. The core workflow—defining requirements, generating a plan, and repeatedly pulling the lever on agent instances—increases distance between the developer and the code, reducing the friction necessary for deep learning.

Key Trade-offs

• Increased system complexity: To mitigate AI's non-determinism, surrounding systems become more complex.
• Skill atrophy: Developers lose critical thinking and coding skills as they shift from writing code to reviewing generated output. Even senior engineers with a decade of experience report brain fog.
• Vendor lock-in: Outages of tools like Claude Code can halt entire teams.
• Fluctuating costs: Token-based pricing is a moving target, unlike fixed employee salaries.

Not Just Another Abstraction

Faye refutes the common argument that agentic coding is just a higher level of abstraction, like moving from assembly to FORTRAN. Unlike past abstractions, today's tools have already shown measurable negative impacts—junior developers lose 50% of the learning process when code generation replaces hands-on problem-solving and debugging. The risk is that without years of friction and direct coding experience, the next generation of senior engineers will never develop the deep understanding required for architectural decisions.

Practical Implications

Successful use of agentic coding requires a developer who can critically review thousands of lines of generated code—a skill that is itself being eroded by the tools. Faye calls for vigilance, emphasizing that the cognitive clarity needed for high-level work is undermined when developers abdicate the friction of writing, debugging, and problem-solving.