Kim et al.
Agent harness optimization mostly improves execution reliability, not task reach
Using pass@k curves, the authors show that apparent gains over stronger baselines can reverse at larger attempt budgets and hide shortfalls on hard tasks.
Top University
Younghyun Kim · Wooseok Seo · Minseo Kim · Seunghyuk Oh · Minjae Lee · Donghoon Kim · +3 more
FuriosaAI · Yonsei University · University of California, Berkeley
Research Digest··2 min read
The authors systematically compare language model agents before and after harness optimization, which adjusts prompts, context, and execution logic without changing weights.
§