Harness² iterates on the harness around a fixed model

Google Research released Harness², an Apache-2.0 framework for changing the system around a model that enables it to act as an agent, rather than its model weights. It exposes prompts, guardrails, memory, tool guidance, subagents, skills, scripts and plugins; its task-level Propose–Probe–Compose step refines a harness on a task, while a separate level retains lessons for later tasks. Why it matters: It makes the optimization boundary inspectable: a team can ask whether a change belongs in the model, the tool loop, memory, or a guardrail. The repository reports benchmark gains, but those are author claims. Its documented benchmark setup also leaves executor network access unrestricted and runs Codex without a sandbox, so review those controls before treating it as a safe evaluation path.