Reproduce “CoWork-X: Experience-Optimized Co-Evolution for Multi-Agent Collaboration System” 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

Zexin Lin, Jiachen Yu, Haoyang Zhang, Yuzhao Li, Zhonghang Li, Yujiu Yang, Junjie Wang, Xiaoqiang Ji

CoWork-X introduces a framework for multi-agent collaboration that enables agents to continuously refine shared skills across episodes while maintaining real-time responsiveness. This system uses an Execute–Optimize loop to decouple fast in-episode execution via HTN-based skills from slower, LLM-driven post-episode optimization, achieving cumulative performance gains with reduced online latency.

Key insights

  • Decoupling fast, rule-based in-episode execution from slower, explicit, post-episode LLM reasoning effectively resolves the real-time coordination versus multi-episode adaptation trade-off in multi-agent systems.
  • Representing agent skills as structured, HTN-executable code enables verifiable, systematic refinement and cumulative learning, moving beyond unstructured textual reflections or prompt engineering.
  • An active co-evolution framework with a shared, iteratively refined skill library facilitates symmetric, peer-to-peer optimization, enhancing the collective behavior of all agents in a collaborative system.

Summary by alphaXiv.

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