Thoughts on Slowing the Fuck Down

6 hours ago

Coding agents have enabled rapid project development but often at the cost of code quality and maintainability.
Software reliability is declining, with frequent outages and UI bugs becoming more common, possibly accelerated by AI-generated code.
Companies using AI to write 100% of their code often produce low-quality products with memory leaks, UI glitches, and crashes.
Agents lack the ability to learn from mistakes, leading to repeated errors and compounding issues in codebases.
Without human oversight, agents introduce excessive complexity and poor architectural decisions, creating unmaintainable codebases quickly.
Agentic search has low recall, making it difficult for agents to refactor or improve large, complex codebases effectively.
Effective use of agents involves delegating narrow, non-critical tasks while maintaining human review and quality control.
Slowing down development allows for better decision-making, fewer unnecessary features, and more maintainable systems.
Human discipline and agency are essential to ensure that AI-assisted development remains productive and sustainable.

Hasty Briefsbeta