When AI coding agents produce slop: the architecture lesson from Dioxus

Jonathan Kelley describes how Dioxus Labs built a cross-platform app framework in Rust from scratch, handcrafting every line until AI coding agents became viable for Rust. The team initially produced thousands of agent-written lines, but little passed their quality bar—they became a 'slop

cannon.' After reflection, they found agents excel at knowledge-heavy tasks like implementing Kotlin/Swift plugins in weeks instead of years, accelerating CSS debugging by knowing browser specs instantly. The key tradeoff: agents enable faster feature delivery but require strong architectural substrate. Poor architecture

produces faster spaghetti. The team now spends most development time on architecture planning rather than code writing. They still review every PR line by line and find prompt engineering surprisingly influential on output quality. Agents freed the small team from mundane verification

tasks, enabling multiple patch releases per week they'd previously feared. However, automated test writing remains weak—agents write correct but shallow tests, missing edge cases. The team still enumerates test conditions manually, though agents excel at building fuzzing harnesses for production-grade quality assurance.