Harness^2: Recursive Agent Harnessing for an Open World
The paper proposes recursive edits to an agent’s execution harness, not its base model, as the path to better open-world performance.
Harness^2 uses task-level recursion to revise a harness for the current task without relying on ground-truth verifiers. A second, harnessing-level recursion extracts lessons from those edits so the agent can adapt better on later tasks. The authors report gains across three professional benchmarks and four model-harness setups, including up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB. Source: HF Daily Papers' note.
Harness^2 uses task-level recursion to revise a harness for the current task without relying on ground-truth verifiers. A second, harnessing-level recursion extracts lessons from those edits so the agent can adapt better on later tasks. The authors report gains across three professional benchmarks and four model-harness setups, including up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB. Source: HF Daily Papers' note.
score 5