Harness² lets agents refine their own setup
Google’s open-source Harness² framework adjusts prompts, tools, memory and other harness components without changing model weights. It tests edits against task behavior, then carries evidence-backed lessons into later tasks; the project supports parallel or sequential refinement with OpenCode or Codex. Its repository reports gains on three professional benchmarks, including up to 11.5 points on WorkBuddy-Bench. A Linux demo needs Python 3.12 and uv; model-backed runs require setup and incur API costs.
Try the no-credentials demo first; restrict network access before benchmark evaluation to avoid exposure to public rubrics.