AI: Building Ambitious Software

Jonathan Kelley observed that this skill gap made it "way too easy to become what we call a slop cannon," producing a large volume of low-quality, inefficient code. The initial failures highlighted that while AI tools offer immense potential, their effective integration requires users to develop new competencies to wield them properly.

AI coding agents are proving instrumental in making notoriously difficult programming languages, such as Rust, more accessible. Jonathan Kelley noted that these agents "deal with the development burden for you" by handling complex aspects like edge cases and the borrow checker. This automation significantly reduces the cognitive load associated with writing Rust applications, effectively eliminating the steep learning curve that once deterred many developers. By abstracting away much of the language's inherent complexity, AI empowers more engineers to engage with powerful, high-performance languages.

Jonathan Kelley stressed that many developers do not adequately consider the ongoing need to build easily maintainable code, which directly impacts software velocity. Achieving this requires consistent maintenance of documentation, tests, and benchmarks, alongside carefully managing the tension between rapidly shipping new features and avoiding API-breaking changes. The quality of the underlying code substrate is paramount for sustained progress and evolution.

Many challenges in complex software development, such as those encountered in Dioxus, are fundamentally "knowledge problems" that no single engineer can fully master. Jonathan Kelley explained that it is not feasible for a team to know every detail about every build system, runtime, operating system, programming language, API, and quirk. This is precisely where coding agents excel, leveraging their ability to process extensive documentation, reverse-engineer APIs, and even dig into binaries with a level of patience far exceeding human developers. AI bridges these knowledge gaps, acting as an omnipresent expert.

Jonathan Kelley cited an instance where Kotlin and Swift plugins for Dioxus were shipped in two to three weeks with coding agents, a process that historically would have taken years with manual development. The core implementation work was completed in roughly a single day, with the remaining time dedicated to building comprehensive test cases and thorough testing on real devices. Agents also prove effective in accelerating debugging for intricate styling and layout issues, further boosting efficiency.

AI's integration into development workflows fundamentally shifts the quality bar, empowering teams to prioritize robust solutions over quick fixes. Jonathan Kelley noted that humans often resort to shortcuts due to project complexity or time constraints, resulting in 'hacky' code. However, with AI agents handling much of the heavy lifting, teams can now invest the necessary time to implement solutions "the right way," fostering a culture of quality. This newfound ability reduces the temptation to compromise standards for the sake of speed.

Jonathan Kelley explained that the core Dioxus team comprises only three engineers, making any time spent on repetitive tasks, such as verifying tarball directory structures, a diversion from critical architectural thinking. AI assistance maintains quality across large codebases and multiple releases, protecting against broken integrations with third-party tools. This automation allows the small team to focus their valuable intellectual capital on solving the software's hard problems.

AI agents significantly improve both code documentation and release velocity, enabling more frequent and confident software updates. Jonathan Kelley highlighted that agents efficiently handle release checklists and bug backports, which are typically time-consuming tasks. Furthermore, AI ensures documentation consistency by keeping comments aligned with code changes, a critical factor for maintainability. This augmented workflow has allowed Dioxus to achieve a weekly or even multi-weekly release cadence, a frequency that would have been intimidating and risky without AI assistance.

For foundational software like Dioxus, complex end-to-end scenarios remain a challenge for agents to effectively test. While AI excels at brainstorming test ideas, its ability to execute sophisticated test strategies is still developing. However, AI demonstrates high utility in specific areas such as fuzzing and building test harnesses, augmenting human efforts rather than replacing them entirely. Human developers are still crucial for designing comprehensive and insightful test strategies that cover critical, non-obvious failure modes and system interactions.

Jonathan Kelley noted that AI can quickly generate code, but it lacks the long-term architectural foresight necessary for sustainable development. Excessive reliance on agents without strong underlying architectural planning can lead to 'spaghetti code' that is difficult to maintain and evolve. Consequently, the majority of human development time is now dedicated to architectural thinking, focusing on future features and how the system will evolve to accommodate them. Human engineers must design robust foundations that AI can then populate with code.

Jonathan Kelley emphasized that "nothing really has changed. Reading code has always been more important than writing code." For open-source projects, line-by-line review of pull requests is still mandatory to uphold high standards. Prompt engineering is increasingly vital to bridge the gap between a user's high-level intent and the AI model's implementation. Ultimately, the core responsibility of a software engineer persists: building elegant solutions and anticipating the long-term evolution of the system, with AI serving as a powerful assistant rather than a replacement for human oversight.