Building ambitious software — Jonathan Kelley, Dioxus Labs & Cognition
Dioxus found that generating more Rust was easy but merging it was not. Agents paid off on bounded research and maintenance work when architecture, real-device tests, and line-by-line review stayed human-owned.
After generating **tens of thousands of lines of Rust**, the Dioxus team found that very little met its merge bar. A Kotlin and Swift plugin implementation arrived quickly, but the team spent **two to three weeks** building test cases and validating it on real devices.
Use agents where patience and broad technical recall matter: documentation research, debugging specifications, release checklists, backports, and documentation consistency. Keep architecture explicit, communicate intent precisely, and review shipped code line by line.
After generating **tens of thousands of lines of Rust**, the Dioxus team found that very little met its merge bar. A Kotlin and Swift plugin implementation arrived quickly, but the team spent **two to three weeks** building test cases and validating it on real devices. Use agents where patience and broad technical recall matter: documentation research, debugging specifications, release checklists, backports, and documentation consistency. Keep architecture explicit, communicate intent precisely, and review shipped code line by line. Test generation remains a weak point: agents can cover an API while missing the behavior that matters. The experience is from one foundational open-source project, so its workflow may not transfer directly, but its core constraint is clear: **code is cheap; quality is not**.
This replaces generation volume with mergeability as the meaningful measure of coding-agent value. Against the candidates, it reinforces bounded tasks, explicit architecture, qualified behavioral tests, and reviewable evidence, while resisting any general move away from line-level review for foundational code. The real-device validation cost also shows that fast implementation can merely relocate effort into verification rather than reduce it.