Harness² Tunes Agent Harnesses, Keeps Weights Frozen

The compelling idea is that agent quality may improve faster by optimizing the surrounding system than by retraining the model. The results are promising, but the extra exploration cost and noisy self-evaluation make production safeguards essential.

• –Task-level recursion compares candidate harness changes without requiring ground-truth feedback.
• –Harnessing-level recursion preserves useful lessons across future tasks.
• –The framework edits eight components, including system prompts, guardrails, memory, subagents, skills, scripts, and plugins.
• –Researchers report gains of up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB.
• –More probing means more tokens, latency, and attack surface; editable guardrails should not be treated as a security boundary.