Turbo Harness: Instance-Adaptive Harness Optimization
Turbo Harness adapts a global agent harness per task instance and outperforms prior optimizers on seven benchmarks.
Turbo Harness adapts a globally optimized agent harness to individual task instances instead of applying one harness everywhere. It recycles artifacts from a completed global optimization into a structured playbook, then trains a harness editor to propose instance-specific patches. At inference, the editor uses the instance and playbook to build a tailored harness for the execution model. Experiments show consistent gains over existing harness optimizers on seven benchmarks covering interactive agents, software engineering, and long-horizon terminal tasks.
- Turbo Harness builds instance-specific patches from a global harness.
- It recycles prior optimization artifacts into a structured playbook.
- A trained editor applies the playbook at inference time.
- It beat baselines on seven agent, coding, and terminal benchmarks.
From the original article
1 min read at arxiv.orgAutomating the search for effective harnesses is an important step toward enabling agents to recursively self-improve. Existing harness optimizations typically produce a single global harness that is applied uniformly across task instances. However, a harness that works well on average may not be optimal for every instance.
By Tunyu Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.