Recursive self-improvement becomes a credible research-agent frontier

An eight-day autonomous loop reportedly rewrote its research-agent code, found seven transferable improvements, and reduced reward hacking; PrimeScientists reportedly improved reward by 10.3% while using 50.6% fewer attempts. Broader claims that AI is completing research work remain unverified, and harness drift, hidden evaluations, limited disclosure, and narrow task distributions leave genuine capability improvement unresolved.