Code execution as a tool

Ablation lab: a 30,000-row file should not be pasted into the prompt. The model writes code, the harness runs it outside the prompt, and only what the task needs comes back, with a note saying where the full file is. Measure the cut on a real file in the browser (99.96% fewer estimated tokens), switch each rule off to see which task breaks, catch a wrong answer with a separate recompute, and read what the published source does and does not measure.

This Harness Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.