Turbo Harness: Instance-Adaptive Harness Optimization
Turbo Harness edits a globally optimized agent harness per task instance, using a playbook distilled from the original optimization run.
The paper argues that one harness applied across all instances leaves performance on the table. Its editor takes the instance plus prior optimization artifacts and generates patches to the global harness before execution. The authors report consistent gains over harness-optimization baselines on seven benchmarks covering interactive agent work, software engineering, and long-horizon terminal tasks. ArXiv · AI/CL/LG's note
The paper argues that one harness applied across all instances leaves performance on the table. Its editor takes the instance plus prior optimization artifacts and generates patches to the global harness before execution. The authors report consistent gains over harness-optimization baselines on seven benchmarks covering interactive agent work, software engineering, and long-horizon terminal tasks. ArXiv · AI/CL/LG's note
score 4