Reproduce “Harness^2: Recursive Agent Harnessing for an Open World” 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
Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu, Bhavana Dalvi Mishra, Yifei Ming, Han Yu, Rujun Han, Chen-Yu Lee and 1 more
Harness² improves an agent’s editable task harness while preserving supported lessons for later edits, without changing the base model. It compares matched runs to identify behavioral changes, then selectively composes harness edits and updates an improvement library after each task. For GLM-5 under OpenCode, its WorkBuddy-Bench macro-average was 75.7 versus 64.2 for the base harness; contrasts do not verify correctness.
Key insights
- Comparing runs under matched conditions can provide useful signals about whether an edit addresses a diagnosed failure, even when no ground-truth answer is available.
- The refinement process itself can yield reusable experience: retaining the reasons and evidence behind edits is different from blindly transferring a task-specific execution harness.
- Broader exploration helps: parallel recursion with five refinement candidates produced further gains over a single refinement step, though it requires additional model calls and compute.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.