Google Research open-sourced Harness², a framework that lets AI agents refine the execution harness around them without changing or training the underlying model. According to figures cited by the supplied report from the Harness² paper, reorganizing system prompts, skills, scripts and subagents raised GLM-5 with OpenCode from 64.2 to 75.7 on WorkBuddy-Bench, while Gemini 3.5 Flash with OpenCode rose from 31.5 to 36.0 on JobBench. The framework uses Propose–Probe–Compose: it drafts separate workflow and domain candidates, reruns them on the same task to compare behavior, and selectively combines or retains the current harness. A second layer reviews accepted changes, stores evidence-backed experience in an improvement library and reuses it in later tasks. The report says behavior comparisons can reveal process differences and regressions but cannot reliably establish task correctness, and that improvements may be task-specific and costly to validate in other settings.