Harness²
Google Research released Harness², an open-source framework that recursively adapts an AI agent's prompts, tools, and orchestration around frozen model weights without ground-truth verifiers.
Rather than fine-tuning models, Harness² optimizes the surrounding agent harness during deployment. It uses a task-level contrastive Propose-Probe-Compose step that guides harness edits without ground-truth verifiers, paired with harnessing-level recursion that extracts experience from these edits into a continual-improvement library for future tasks.
- Date
- Friday 9 October 2026
- Lab
- Kind
- infra
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| WorkBuddy-Bench score improvement | +11.5 points Best single-step gain over base harness across GLM-5 and Gemini 3.5 Flash configurations | company |
| LAB accuracy improvement | +11.8 percentage points Best single-step gain over base harness across GLM-5 and Gemini 3.5 Flash configurations | company |
| JobBench score improvement | +4.5 points Best single-step gain over base harness across GLM-5 and Gemini 3.5 Flash configurations | company |
Code is released under Apache 2.0 at github.com/google-research/harness2; marked as a research release rather than an official Google product.
Sources
- arxiv.org/abs/2610.14224
- pasqualepillitteri.it/en/news/25395/google-harness2-ai-agents-improve-themselves
This record was checked and corrected against its sources on 18 October 2026. How we check