Reproduce “Exo: An Architecture for Self-Improving Systems at Runtime” 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
Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica, Martin Casado, Ankur Goyal
Exo lets agents edit their harness policies during task execution while keeping model weights fixed. Agents can revise tools, context construction, and improvement strategies, with a protected runtime retaining execution history and controlling policy replacement. On AppWorld, held-out model cost averaged $0.141 per task versus $0.359 for a no-edit control, with equal 96.9% pass rates; this evidence is benchmark-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.