AI News

Harness² is a plug-and-play framework for improving the execution harnesses that determine how large language model agents operate. The paper targets open-world settings, where tasks are unseen during harness design and test-time edits lack reliable ground-truth verifiers. Its task-level recursion uses a contrastive Propose-Probe-Compose procedure to guide harness changes without such verifiers, with additional parallel or sequential recursion available to expand the search. A second, harnessing-level recursion extracts experience from those edits and updates a continual-improvement harness so that later tasks can benefit. Across three professional benchmarks and four model-harness configurations, one recursive step consistently improved GLM-5 and Gemini 3.5 Flash over their base harnesses. Reported gains reached 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. Increasing the task-level recursion budget produced further improvements, according to the abstract.