Google Research open-sources Harness², a framework for agents that improve their own prompts and tools at test time
Per Google Research's own GitHub release and accompanying arXiv paper, a "harness" is everything wrapped around a model — prompts, tools, memory, skills, orchestration — and Harness² refines that surrounding system during deployment while keeping model weights completely frozen. It uses a "Propose, Probe, Compose" loop to improve behavior through behavioral contrasts rather than labeled examples, keeps a "lessons library" that carries improvements forward to future tasks, and already supports OpenCode and Codex as executors. It's Apache 2.0 licensed and has picked up over 11,000 GitHub stars within two days.
Signal: a real research contribution with public benchmarks, not just a demo — a single Harness² step lifted GLM-5 by up to 11.5 points on WorkBuddy-Bench and Gemini 3.5 Flash similarly on other benchmarks, and it works without labeled data or ground-truth verifiers, which the team says most self-improvement methods require.