🔗 Thoughts on slowing the fuck down

Coding agents have become increasingly prevalent in production codebases over the past year, but their widespread adoption is creating brittle, error-prone software systems. The article observes that software quality is deteriorating across the industry, with major companies experiencing outages and quality issues that correlate with increased AI-generated code.

Agents lack human learning mechanisms and create compounding errors without bottlenecks to slow accumulation. When agents operate autonomously without human oversight, small code quality issues multiply into unmanageable complexity within weeks, creating codebases that even agents cannot effectively refactor or improve.

Effective agent use requires scoped tasks with clear evaluation functions and human oversight at quality gates. The article advocates for using agents for boring, non-critical work while humans retain control over architectural decisions and final implementation details.