Rohan Paul (@rohanpaul_ai)T2·Agents
78 pts
A Northwestern and Google paper introduces Harness2, which lets agents self-improve their harness (prompts, tools, skills, memory) on new tasks without answer keys by test-running proposed edits and keeping only those that help. A single improvement step lifted GLM-5 by 11.5 points on WorkBuddy-Bench, surpassing Gemini 3.5 Flash on its stock harness, while unvalidated harness edits dropped coding scores from 73.4 to 63.2.