Agent harnesses aren't dying: harness distillation changes the job, says O'Reilly

bibryam · x · 2026-11-05

Laurie Voss (O'Reilly/Arize) responds to "are agent harnesses dying?" Answer: no, but the role is shifting. Trigger: Peking University's Harness-Zero paper showing harness distillation—training a specialized harness's behavior into a model—beats the original model+harness combo. Core argument: generalizable harness pieces get absorbed into models; what remains should only carry your tools, data, users, and environment, so plan to regularly discard and rebuild harnesses. The article ships a concrete loop for each new model release: tag every trace with harness/model versions, log each intervention, disable harness components one by one on the new model, rerun evals and keep only what still helps—reading traces, since scores alone hide which change moved results.

Original post →

More from coding & agent

coding & agent channel →