Adding Friction: Why You Should Work Harder Than Your AI

Relying on AI agents for code generation can degrade a developer's skills by bypassing the cognitive effort required for learning. The author suggests reintroducing 'friction' through manual coding, deep questioning, and delayed AI usage to maintain mental engagement. Ultimately, the goal is to ensure the human remains the primary thinker, even if it reduces short-term speed.

Key Points

• Agentic code generation bypasses the brain's synthesis processes, leading to poor skill retention and cognitive 'brain fog'.
• The instant reward system of AI tools mimics addictive social media feeds, discouraging the deliberate effort needed for deep learning.
• Adding intentional friction, such as writing code manually before asking for an AI review, helps solidify a developer's technical foundation.
• Using AI for questioning, critiquing, and exploring documentation is more beneficial for long-term growth than simple code generation.
• Short-term speed gains from LLMs should not come at the expense of the developer's own cognitive engagement and expertise.

Sentiment

The overall sentiment is cautiously supportive of the article's warning. Most commenters recognize the risk of shallow understanding and agree that developers need to add friction through active review, refactoring, tests, and critique, but the thread is not anti-AI. The dominant view is that agentic coding is useful only when the human remains the more disciplined, mentally engaged participant. Opposition comes mainly from people who see AI as a higher-level tool or who believe rigorous engineering process prevents deskilling, not from people denying the need for human judgment.

In Agreement

• Passive agentic coding can leave developers with code they do not understand or remember, because the rewarding part of getting a working answer replaces the slower process of building a mental model.
• Effective AI use should make the human mentally engaged through refactoring, test review, documentation checks, plan critique, and deliberate pushback rather than simple acceptance of generated patches.
• The fatigue many developers feel after AI-assisted coding may be a sign that the real work has shifted from typing to supervision, judgment, and continuous verification.
• LLMs can be useful for learning when they provide exercises, explanations, or testable claims, but using them as answer machines risks weakening retention.
• Businesses that optimize only for feature churn may underprice the long-term cost of code that no one on the team deeply understands.

Opposed

• Many refactoring tasks described as agent work are already handled faster and more reliably by deterministic IDE, LSP, structural-search, or language-aware tooling.
• Experienced developers can avoid deskilling by treating agents like junior collaborators and maintaining rigorous specifications, interfaces, test loops, reviews, and documentation.
• AI coding may be another abstraction layer rather than a degradation of craft, shifting value from low-level implementation details toward higher-level design and algorithmic judgment.
• For non-engineers or people who could not previously build software, AI tools create access and help develop a different kind of judgment rather than replacing an existing skill base.
• Avoiding LLMs entirely may preserve craft, but some commenters argue that competitive work environments will still demand the productivity benefits of automation.