Reproduce “Self-Healing Harness for Runtime Oversight of Agent Self-Modification” 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
Sina Tayebati, Divake Kumar, Nastaran Darabi, Ranganath Krishnan, Amit Ranjan Trivedi
The Harness lets an agent propose behavioral rules while an external runtime decides which become persistent. It evaluates traces, tests candidate rules through replay or forward trials, and requires measured improvement without protected-case regression when replay is available. Among 383 replay-decided retirements, 211 improved the triggering failure but degraded protected behavior. This identifies conflicts, not broad safety or alignment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.