Slowing Down: Why Human Agency is Essential in the Age of AI Coding

The rise of AI coding agents is leading to a crisis of software quality where speed is prioritized over maintainability and correctness. Because agents lack a global view of codebases, they introduce compounding errors and unnecessary complexity that eventually make systems unmanageable. The author advocates for a disciplined approach where humans manually handle architecture and use AI only for scoped tasks they can thoroughly review.

Key Points

• AI agents lack a natural bottleneck, allowing small architectural errors to compound into unmanageable technical debt much faster than human developers.
• Agentic search has low recall in large codebases, leading to code duplication, inconsistencies, and 'learned complexity' from training data.
• The 'magic' of high-speed code generation often results in 'slop'—software filled with memory leaks, UI glitches, and broken features that lack human oversight.
• Maintaining friction in the development process is necessary for learning and for understanding the 'gestalt' or overall health of a system.
• Developers should limit AI output to a volume they can realistically review and keep human agency at the center of critical decision-making.

Sentiment

The community overwhelmingly agrees with the article's core thesis. Most commenters share the view that human oversight, understanding, and craftsmanship are essential when working with AI coding agents, and that uncritical adoption leads to unmaintainable codebases and deskilling. The minority opposing view focuses on AI's genuine utility for small tasks and the likelihood that costs will decrease, but even these commenters generally acknowledge the need for human discipline. The overall tone is cautious and experience-driven rather than hostile or dismissive.

In Agreement

• The act of writing code builds irreplaceable tacit knowledge and mental models that cannot be gained from merely reviewing AI-generated diffs
• AI agents are 'merchants of complexity' that produce verbose, sprawling code where concise solutions would suffice, and developers must enforce discipline through scoped tasks and strict review
• Companies that let only AI understand their codebases create dangerous vendor lock-in and lose the ability to debug, maintain, or operate their own systems
• The real productivity bottleneck is human comprehension, not code generation speed, and exceeding that limit leads to burnout or indifference to accumulating errors
• Drawing from DevOps and Toyota's Andon cord philosophy, cultural discipline and willingness to stop and fix root causes is more important than shipping speed
• Current AI hype mirrors previous technology cycles like NoSQL and microservices, and will likely settle into a narrower, more appropriate niche over time

Opposed

• AI inference costs will continue falling as hardware improves, making the vendor lock-in and price-gouging fears overblown, similar to how ridesharing and cloud computing became sustainable businesses
• Vibe coding is genuinely useful for small, bespoke tools and internal business applications, similar to how Excel solves specific problems without needing to be maintainable
• AI agents democratize software creation for non-programmers, allowing artists and domain experts to build things that would have been impossible before, which has real value even if the code quality is low
• Disciplined use of agents with TDD, skills files, and specification-driven development can produce reliable results even in unfamiliar programming languages
• Open-source models and local inference will prevent monopolistic pricing, as the AI market has competitive dynamics unlike OPEC's control of oil supply