Reproduce “SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations” 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

Xinping Lei, Minghao Liu, Ruijie Zhu, Xiaohui Hu, Xingqiao Li, Ao Zhang, Wenhao Chai, Han Li and 18 more

SWE-CoWork evaluates whether coding agents can recover workplace requirements and deliver verified changes from brief handoffs. Across 101 assignments, agents consult simulated colleagues and records before implementing and testing code. GPT-6 Astra passes 62.91% of new-behavior test nodes, but fully passes 7.92% of tasks; giving it complete requirements raises the node pass rate to 78.84% without changing task success.

Summary by alphaXiv.

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