Google Cloud AI Research and Northwestern University propose Harness², a way to improve an agent's prompts and tools while it works, without changing the model

Agents also rely on hand-tuned prompts and tools, and new tasks have no answer key. The team proposes Harness², which keeps the model fixed, tests a few changes on each task and keeps what helps. The paper reports it lifted GLM-5's WorkBuddy-Bench average by 11.5 points over a plain OpenCode setup. But extra runs cost more compute, and with no answer key the changes may be wrong or not transfer.

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.