Self-improving Agent Harnesses Overfit. How Google Fixes It

StackSweep · 3 weeks ago · 27

Google Cloud AI Research found that letting an agent evolve its own prompt/tool harness overfits the evolve-set benchmarks and often loses generalization on held-out benchmarks. Unregularized evolution reached 92.8 on the evolve set but only 40.3 on held-out benchmarks, while its RRSI regularized method reported 90.5 on the evolve set and 43.6 held-out using 2.42M vs 3.80M policy tokens per trial. The fix changes harness evolution by regularizing the search with edit budgeting, noise-aware acceptance, a leakage critic, and pruning, which improves out-of-distribution gains without reverting to memorization.