Google Research has open-sourced Harness2, a framework that improves an AI agent by rewriting the harness around the model at test time. The harness is the code and scaffolding that connects a model to tools, observations, and actions; Harness2 treats that machinery as something to be recursively refined rather than fixed before deployment.
According to MarkTechPost, one refinement step lifted GLM-5 by 11.5 points on WorkBuddy-Bench without any weight updates. That suggests agent performance can be improved by optimizing how the model is controlled and called, not only by training a better model.
The source is a single announcement, so there is no independent comparison to check against. Still, the Apache-2.0 release makes Harness2 a concrete, permissively licensed example of recursive test-time harness optimization.