Reproduce “Evolving Agent Teams” 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
Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica
Evolving Agent Teams changes how coding agents are organized during GPU kernel optimization. A supervisor diagnoses stalled progress, redesigns roles and coordination, and preserves lessons and code in a shared filesystem. On DeepSeek-V3 MLA paged decode, additional-budget runs achieve 1.21× geometric-mean speedup over production FlashInfer across 47 H100 workloads, with 36 faster; 11 remain slower, without beating the run-stage-only reference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.