Harness²: Recursive Agent Harnessing for an Open World (Google Research) (arxiv.org)

Google Research has announced Harness², a pioneering framework that enhances the capabilities of large language model (LLM) agents by addressing critical challenges in deploying recursive self-improvement (RSI) in open-world scenarios. The twofold approach of Harness² employs task-level recursion and harnessing-level recursion to refine agent performance. Task-level recursion utilizes a novel Propose-Probe-Compose method to enhance the harness without needing ground-truth verifiers, allowing for effective online adjustments even in unseen tasks. Meanwhile, harnessing-level recursion builds a continual improvement system that draws on the insights gained from previous edits, enabling agents to adapt more proficiently to future challenges. This development is significant for the AI/ML community as it not only demonstrates a robust method for enhancing LLMs' adaptability but also sets a precedent for handling the complexities of open-world tasks. Experimental results show that applying Harness² leads to substantial performance improvements for models like GLM-5 and Gemini 3.5 Flash, achieving gains of up to 11.5 points on WorkBuddy-Bench and notable advancements on other benchmarks. By facilitating a scalable approach to recursive agent harnessing, Harness² could greatly influence the future of AI model training and deployment in dynamic environments.