What's this about?
John Gruber highlighted Claude Code creator Boris Cherny describing an experiment: having Claude Code rewrite the Electron-based Claude desktop app in native Swift, with a verification loop — "run the app in a Mac VM, screenshot it, compare it to the Swift version pixel by pixel."
What exactly is "Electron vs. native"?
Electron apps are websites wrapped in a desktop window — one codebase everywhere, but heavier and less Mac-like. Native Swift apps are built with Apple's own tools: faster, lighter, platform-correct. Rewriting Electron→native is expensive drudgery humans rarely fund, which is why "let the agent do it, and let it verify itself by comparing screenshots" is an interesting test case.
What's the community saying?
HN 36 pts, 9 comments
- Broad agreement with Cherny's core claim — "verification is the single most important thing people get wrong" with agents.
- But pushback that "compare pixel by pixel" is naive: pixel-identical is over-specific, hard, and counterproductive as an acceptance test.
- Cost skepticism: two weeks of tokens could be tens of thousands of dollars — why not hire someone who knows Swift?
- A bigger-picture note: agents now write code humans throw away, and cheap "write one to throw away" exploration may be the actual win.