Blog
What we measured while building this. Numbers we took ourselves, with the method next to them, and the failures included. A claim without a way to check it is just marketing with a decimal point.
The number we kept quoting about Electron was the wrong one
Swapping Electron for Tauri took the shell from 221 MB to 117, and later to 90. The figure we had been repeating was 627 against 117, which compares an entire application against one part of one, and we were the ones repeating it.
Measuring the app at rest told us nothing
Twelve tabs open, no long tasks, around 80 fps. The app looked healthy at rest and switching tabs took half a second. Freezing hidden panes cut the median switch from 491 ms to 352, and a second run on a busier machine only got 8%.
A commit about a landing page deleted two React components
Two agents shared one checkout, so they shared one git index. A commit that touched only the marketing demo swept in another agent's staged deletions and left main importing two files that no longer existed, for forty-one minutes.
Eight panes, and the sidebar animation costs two frames each
Animating a sidebar by changing padding costs 8 ms of layout per frame with two panes open and 66 ms with sixteen. The frame budget at 60 fps is 16.7 ms. Doing the same movement with a transform costs zero at every count.
We sent 244 tokens for every token we got back
Three months of building this app with coding agents, read out of its own database: 2.05 billion input tokens, 8.4 million output, and 9,870 dollars at API rates. Four fifths of the input was never cached, and that is almost the whole bill.
The app was holding 14 GB, and reloading it made things worse
Sixty-five WebContent processes, sixty-one of which used no CPU at all over twelve seconds, holding 11.7 GB between them. Every reload created more of them, and the reload was the first thing anyone suggests.
Our product screenshots were 3× too small, and we could prove it
One of our six product images was rendering the app's own 13px interface text at 4.7 effective pixels. Here is the ratio that catches it, what the sites that get this right actually do instead, and the check that now fails the build.