Harness2 (github.com)

Harness² has been introduced as a plug-and-play framework designed for recursive agent harnessing in open-world environments where agents face unseen tasks during the harness design phase. This innovative system consists of two levels of recursion: task-level and harnessing-level. The task-level focuses on refining the execution harness for specific tasks using a contrastive Propose–Probe–Compose step, allowing for adjustments based on behavioral contrasts without relying on ground-truth feedback. Meanwhile, harnessing-level recursion develops a persistent improvement harness, leveraging insights gained from previous tasks to inform future adjustments. The significance of Harness² for the AI/ML community lies in its substantial performance improvements across various professional benchmarks, enhancing models like GLM-5 and Gemini 3.5 Flash by as much as 11.8 percentage points on task completion rates. By supporting both parallel and sequential recursion modes, along with a variety of customizable components, Harness² empowers developers to optimize agent performance effectively. These features represent a significant step towards more adaptable, self-improving AI systems capable of handling complex tasks in dynamic environments, ultimately driving innovation in the field of artificial intelligence.