Notes · measured, losses included

Write-ups from the bench

What breaks in real agent harnesses, measured on real repos, with the numbers published whether or not they flatter my own tools.

2026-08 · benchmark write-up

Your agent stack doesn’t tell you when it’s wrong

A matched-conditions A/B of the popular open-source agent stack against my own, on four real repos, every claim grep-verified. Both stacks failed silently, in opposite directions. The full matrix is in the post, my own losses included.