As reported by finance.biggo.com, Kelley argues that AI coding agents crossed a real threshold on Rust work over the past six months — but that "code is now cheap, but quality is not," and the gap between the two is now where the engineering actually lives. For anyone shipping custom web and business applications, that framing is worth stress-testing against the marketing.
The five-year bet
Kelley traces Dioxus back to 2021, when the pitch was unfashionable: write everything in Rust, use HTML and CSS as markup, get native performance with no JavaScript runtime, no IPC, and no virtual machine. The team committed when React Native felt "janky" and Flutter was "too slow" for tight native API work. Five years later the project reports 37,000 GitHub stars, roughly 200 million cumulative end users, bundle sizes under 5 MB, RAM footprints under 50 MB, and hot-reload patches in the 100-millisecond range. The point of those numbers, as Kelley frames them, is that the entire framework — reactivity, font rendering, hot reloading, bundling — had to be built from scratch, because in 2021 there was essentially nothing off the shelf for that stack. He summed up the original decision as choosing to build "not because it's easy, but because we thought it would be easy."
Where agents actually moved the needle
Over the past six months, Kelley's three-person core team found that AI agents could ship a Kotlin and Swift plugin integration in two to three weeks, with the implementation itself landing in roughly a day and the remainder absorbed by real-device testing. Mundane release chores and test-harness plumbing, which used to consume a significant slice of the team's bandwidth, collapsed into inference calls. The probabilistic shortcuts are real. They are also bounded.
Agents, in his team's experience, can produce thousands of lines of feature code that never clear the merge bar. They write tests for constructors instead of behavior. They leave documentation comments behind the API surface they actually changed. Kelley calls the pattern a "slop cannon" — high volume, low signal — and argues that the review bar, not the typing speed, is what determines what ships.
What this means for the visual-engineering stack
That distinction matters for no-code and low-code platforms more than it first appears. Visual engineering tools have always promised to compress the gap between intent and implementation. If agentic AI collapses the same gap in text-based code, the differentiator migrates upstream — to architecture, to the judgment of what to build, and to the discipline of reading every diff before it lands. Two questions worth keeping open: whether the slop-cannon pattern stabilizes as models improve or deepens as code volume outruns review capacity, and whether platforms built around visual abstraction absorb the same collapse or whether the architectural layer becomes the durable competitive surface. Kelley is betting on the latter. The next six months of tooling releases will be where that bet gets audited.