Google Research released Harness² code, which changes prompts, tools, memory, and orchestration while keeping model weights fixed. Its two-level recursive structure proposes changes for the current task, compares execution differences and synthesizes improvements, then carries adopted configurations and evidence into the next task. It supports Codex and OpenCode runners and reports gains of up to 11.5 points in particular evaluations. It is not an officially supported product, and default network access and execution settings require care. The repository also advises blocking access to public answers and scoring criteria, so performance and cost should be assessed together in controlled environments.