Preprint reports benchmark gains from agents revising their execution workflows

Researchers report that Harness^2, a framework that lets AI agents revise the workflows governing their execution, improved performance across three professional benchmarks and four model-workflow configurations. The preprint reports gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. The framework probes proposed workflow changes without ground-truth verifiers, then draws on those edits to improve adaptation to future tasks. A single revision step consistently improved GLM-5 and Gemini 3.5 Flash over their base workflows, with further gains as the revision budget increased. The reported results concern benchmark tasks rather than demonstrated production deployments.