Reproduce “Harness Engineering for Software Engineering via Modular Executable Dev-Primitives” 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

Haibo Jin, Xinjie Li, Peng Kuang, Haohan Wang

HERMES assigns LLM-backed interfaces to relevant repository artifacts, including source files, tests, and configurations. Components reason about and edit their artifacts, message related components, and use execution evidence to guide revisions. Across four software-engineering benchmarks, it averages a 12.4-point gain over matched baseline harnesses with GPT-5.6 Sol at medium effort; results do not show that context growth is eliminated.

Key insights

  • Component-specific reasoning offers an alternative to making one agent maintain a complete, continually updated model of the repository; directed communication helps convey cross-file constraints.
  • Execution-grounded diagnosis is central to the reported gains: removing diagnosis feedback caused the largest drops among the tested ablations.
  • Stronger models appear most useful for selecting components and diagnosing failures, while smaller models can handle many component-level tasks at lower cost.

Summary by alphaXiv.

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