After two years of experimenting with AI coding agents, a developer is abandoning the practice. The author describes a predictable arc: initial awe at simple tasks, then frustration with complex ones. Agents produce plausible-looking but ultimately incoherent codebases, a phenomenon termed slop. The core issue is a lack of holistic respect for architectural integrity and long-term codebase evolution.

Spec-driven development fails because AI agents can't adapt specifications over time like human engineers. They make rigid upfront decisions and struggle with multi-week projects. The resulting code, while syntactically correct in isolation, becomes a messy, nonsensical whole. This mirrors the experience of vibewriting, where individual chapters seem fine but the novel lacks coherence.

Faced with shipping an unreliable product, the author chose to return to manual coding. This decision was driven by a commitment to user trust and data protection. Surprisingly, this shift resulted in greater speed, accuracy, and overall productivity when accounting for the full cost of debugging and maintaining AI-generated code.