Reproduce “When Sub-Agents Work in Parallel: The Promises and Pitfalls of Dynamic Concurrency in Long-Horizon Coding Tasks” 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
Han Li, HanHaoNing Li, Ziqian Jiang, Yiling Lou
The authors evaluate native dynamic concurrency in three coding-agent systems across five benchmarks, comparing each configuration with and without sub-agent delegation. Results vary: Claude Code’s task-pass rate rises from 7.1% to 21.4% on LoopsBench but falls from 83% to 59% on SWE-bench Verified. Concurrency uses 1.41–3.31 times as many tokens, and mean runtime rises in 14 of 15 agent–benchmark comparisons.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.