Google Research released Harness², a framework for recursive agent harnessing. It enables task-level and harnessing-level recursion to refine agent systems. The framework supports parallel and sequential recursion modes. It works with OpenCode and Codex executors. It is tested on JobBench, LAB, and WorkBuddy-Bench. A single step improved GLM-5 and Gemini 3.5 Flash scores on these benchmarks. It exposes eight editable harness components. The code is open-source under Apache 2.0. A paper describing the work is on arXiv. Installation requires Linux, Python 3.12, and uv. Model-backed runs need additional dependencies and Google Cloud setup.
What it unlocks
Running recursive agent harnessing with task-level and harnessing-level refinement on professional benchmarks.
- 11.5 points gain on WorkBuddy-Bench
- 4.5 points gain on JobBench
- 11.8 percentage points gain on LAB partial pass rate
- 2610.14224 arXiv paper
What you need to act on it
- Linux
- Python 3.12
- uv
- Node.js 20+ for model-backed benchmarks
- Google Cloud CLI for Vertex AI models
- Docker for WorkBuddy-Bench
- Pandoc and Poppler for LAB
Sources