Google open-sources Harness²: agents that rewrite their own harness, lifting Gemini 3.5 Flash by up to 11.8 points
_philschmid · x · 2026-10-19
Google Research released Harness², a plug-and-play recursive framework that lets a frozen model improve its own execution harness on unseen, open-world tasks — addressing the lack of ground-truth verifiers and the task-specificity of prior RSI work.
How it works (four steps)
- The agent runs the base harness per task to record a baseline trajectory
- An Editor agent inspects the trajectory, drafts 2 candidate edits (SI, Skills, hooks, …) and runs both
- A Composer agent diffs new trajectories against the baseline, merges edits that actually fixed failures (within a cap), and runs the final harness
- A Reflector agent logs verified lessons and saves winning components to a library future tasks can pull on demand
Results
- Up to +11.5 on WorkBuddy-Bench, +4.5 on JobBench, +11.8 points on LAB for Gemini 3.5 Flash; with parallel recursion (k=5) it beats the strongest baseline in every domain
- Skills and scripts drive 50% of edits; dropping skills costs 5.3 points
- The model improves on all 3 benchmarks even when editing its own harness; gains scale with the recursion budget
Paper and code: google-research/harness2 on GitHub, harness2.github.io.
More from coding & agent
- Short film storyWeave: Opus wrote the script, effects, and half the transitions — hq4ai · 2026-10-19
- Rewriting Windows 98's Task Manager in Go for $120 in 24 Hours with Claude — derjanni · 2026-10-19
- Opus 5.5 edits video in Claude Code: one prompt, 15 jobs, 100/100 QA score — socialwithaayan · 2026-10-19
- Bare IPython Harness Matches Production Agents on Qwen3.5-9B; Output Normalization Kills Run-to-Run Variance — HauntingMoment · 2026-10-19
- Walking 800m Buys You Half an Apple: an Honest Calorie AI with an Explicit Abstain Layer — Cicada-Objective · 2026-10-19
- Claude Remote Control hit with login deadlock: /login unavailable on mobile — drscotthawley · 2026-10-19