Reproduce “Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yilun Yao, Xinyu Tan, Chao-Hsuan Liu, Yaoming Li, Zhengyang Wang, Zongwei Lv, Wenhan Yu, Zhewen Tan and 5 more

Harness-Bench compares native agent harnesses across models on workflow tasks, grading the artifacts they deliver and recording execution. It gives configurations identical inputs while retaining each harness’s prompts, tools, context management, and recovery. On general tasks, GLM-5.3 reaches 35.77% full credit with OpenCode and 22.26% with Codex, showing pairing matters without identifying which harness feature causes the gap.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.