Reproduce “From Watching to Steering: Enabling Human-Coding Agent Collaboration through Execution Traces” 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

Xiaoke Yan, Jiaying Lei, Danqing Shi, Zichen Wang, Nan Cao

VibeTrace makes coding-agent traces interactive: users inspect actions, fork earlier steps to redirect runs, and convert accepted feedback into reusable skills. Its held-out SWE-bench Verified test showed a 68.0% pass rate versus 65.0% without skills, with lower input tokens, cost, and elapsed time. Feedback came from verifiers, and the frozen skill pool excludes skill-creation costs.

Summary by alphaXiv.

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