Eval engineering explained: turn "correct" into machine-checkable tests

Roger_M_Taylor · x · 2026-09-15

A long-form explainer on eval engineering: the model does the work, but someone must say "this is right" — as long as that's you, your agent can't run without you. Eval engineering means writing down what "right" means so a machine can check it; without it, changing a prompt or swapping a model gives you no way to tell if you broke something.

Method: start from failures — hand-read 100 real failed runs, describe the errors in your own words, and group them; 3-4 failure types usually cover most cases and won't be what you expected. That's why generic judges fail (Hamel Husain's example: a "helpfulness" judge passed an answer connecting a bathroom to the wrong place). Then build checks in two layers: code first, model judges second.

It accompanies a released Eval Engineering Skill that helps coding agents build evals from repo context and agent traces.

Original post →

More from coding & agent

coding & agent channel →