Google has open-sourced Harness², a new framework designed to enhance AI agent capabilities by enabling them to autonomously modify their own code, system prompts, and toolchains during operation. This recursive self-improvement (RSI) approach addresses key challenges in AI development, such as the lack of standardized evaluation metrics and the tendency for models to overfit to specific tasks. Harness² employs a dual-layer recursive architecture: a task-level recursion for experimental modification and a harnessing-level recursion for accumulating and transferring learning experiences across tasks. Initial tests show significant performance boosts on benchmarks like WorkBuddy-Bench, JobBench, and LAB, with some instances surpassing more advanced models. AI
IMPACT This framework could significantly improve AI agent performance and adaptability by enabling them to learn and adapt in real-time, potentially reducing reliance on constant manual prompt engineering and model retraining.
RANK_REASON Google released a research paper and open-sourced a new framework for AI agent self-improvement. [lever_c_demoted from research: ic=1 ai=1.0]
- Codex
- Gemini 3.5 Flash
- GLM-5
- JobBench
- Kaize Ding
- OpenCode
- Ruiyao Xu
- WorkBuddy-Bench
- Yanfei Chen
- Zhongying CuiZhu
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
