Agent harnesses, skill evolution, and RL post-training

Recent work suggests performance depends substantially on harness design and credit assignment, not only the executor model: a Google Cloud AI Research harness reportedly raises AutomationBench from 61.6% to 77.4%, while GRSE targets reusable skills. JOLT reports 6.5× fewer tokens at matched MATH accuracy, with mixed agent-benchmark results. The new AKV serving design adds an inference-systems angle, claiming 27.7% lower costs and 1.5× better P95 latency; independent validation and detailed comparisons remain important.