Recursive Harness Engineering (academy.dair.ai)

Researchers have introduced Harness^2, a groundbreaking framework for recursive agent harnessing in open-world environments. The framework addresses significant challenges in deploying recursive self-improvement (RSI), particularly when tasks are novel and lack established ground-truth verifiers. Harness^2 employs a dual-level recursion strategy: task-level recursion refines the harness through a contrastive Propose-Probe-Compose mechanism, allowing for reliable edits, while harnessing-level recursion utilizes insights from these edits to enhance future adaptability. The significance of Harness^2 lies in its ability to improve the performance of large language models, such as GLM-5 and Gemini 3.5 Flash, across various professional benchmarks. Results demonstrated substantial enhancements, with performance increases of up to 11.5 points on WorkBuddy-Bench and 4.5 points on JobBench. This model not only furthers the capabilities of AI agents in handling unseen tasks but also sets a precedent for future AI/ML frameworks aiming to adapt continuously in dynamic environments, showcasing the potential of scalable recursion in enhancing machine learning harnesses.