Jonathan Kelley uses Dioxus, a cross-platform Rust application framework, to examine what changes when coding agents become capable contributors to ambitious software. He describes an initial burst of generated features that mostly failed the team’s merge-quality threshold. The lesson is that producing more code does not remove the need for maintainable foundations, stable APIs and coherent releases.
Jonathan Kelley identifies knowledge-intensive work as a strong use case: reading unfamiliar documentation, investigating platform APIs and debugging CSS behavior. He reports that agents helped implement Kotlin and Swift integrations quickly, while testing on real devices took substantially longer. The distinction separates initial implementation speed from the work needed to ship confidently.
Jonathan Kelley also describes using agents for repetitive maintenance, including release checklists, backports, documentation consistency and packaging checks. For a small framework team, these tasks matter because external developers depend on the code, examples and interfaces, not just the appearance of a finished application.
Jonathan Kelley gives a mixed assessment of AI-generated tests. He says agents can create superficial assertions without selecting the right behaviors to test, so the team still defines important conditions and test APIs. He finds stronger value in agents helping construct fuzzing harnesses that exercise software with varied and malformed inputs.
Jonathan Kelley argues that architectural judgment and explicit communication become more important as implementation accelerates. The team continues to review every pull request line by line, using AI review as assistance. His conclusion is that lower code-generation costs increase the importance of deciding how a system should evolve and checking what is actually shipped.
Watch on YouTube
