Northwestern and Google Paper Shows Agent Harnesses Can Self-Improve
The non-obvious takeaway is that harness quality, not the underlying model, can decide which agent wins on a benchmark — a weaker model with a self-tuned harness beat a stronger one on its default setup. Developers building agents should treat harness configuration as a per-task artefact to be revised and validated on real runs. Teams without labelled evaluation data benefit most, since the method needs no answer keys, only reruns to filter edits.
