Dioxus maintainer reports tens of thousands of agent-written Rust lines that never cleared the merge bar

Sources: watch, project
Channel: AI Engineer (2026-09-11)
Summary: Jonathan Kelley says the Dioxus team maxed out coding-agent subscriptions and produced tens of thousands of lines of Rust covering long-wanted features, that almost none of it was merged, and that it still sits in draft, a failure mode he names the slop cannon. He is equally specific about where agents paid off: deeply integrated Kotlin and Swift build plugins went from years of hand-written work to weeks, and release checklists, backports and documentation accuracy became tractable for a team of three. He states tests did not, because agents write tests for any API but rarely the right ones, though they do well at fuzzing harnesses, and he treats Rust's learning curve as an advantage now that agents absorb the borrow checker.
Why it matters: A named maintainer putting volume, merge rate and the specific categories that worked against each other is harder to dismiss than either side of the agent-productivity argument in the abstract.