Harness^2 proposes test-time edits to AI agent harnesses without a verifier

DAIR.AI shares a paper it describes as from Google on improving AI agent harnesses for open-world tasks. It says Harness^2 tests several edits for the current task, keeps the best-performing combination and carries lessons into a second harness for future tasks. DAIR.AI reports improvements for GLM-5 and Gemini 3.5 Flash in every configuration tested across WorkBuddy-Bench, JobBench and LAB. Five parallel candidate edits added another 2.9 to 3.9 points on WorkBuddy-Bench, it says.