Harness² FAQ
Last updated: October 19, 2026
GPU needs, fine-tuning, supported models and executors, benchmarks, Windows and macOS, licensing, and Google support status — answered from the official docs.
These answers stick to the official README and documentation. Where the docs are silent, the answer says so instead of filling the gap.
Do I need a GPU to run Harness²?
The README and setup guide do not list a GPU requirement. The credential-free demo runs locally with deterministic editors and real Python executions; model-backed benchmark runs call the configured Vertex AI models.
Is Harness² fine-tuning or SFT?
No. The README says Harness² keeps model weights fixed and optimizes the harness around the model. It is not supervised fine-tuning and does not change model parameters.
Which models does Harness² support?
The official results cover GLM-5 and Gemini 3.5 Flash. The setup guide describes Vertex Gemini executors and judges, and says the improver can be Vertex Claude or Gemini. The docs do not publish a single formal supported-model list, so treat other configurations as unverified until you test them.
Which executors are available?
OpenCode is the default executor. Codex is available through --substrate cx, which requires the optional bridge setup described in the Codex executor guide.
Which benchmarks are included?
JobBench (professional work), LAB (legal work) and WorkBuddy-Bench (office, web and code). Each installs separately with harness2-setup; WorkBuddy also takes --subset office, --subset web or --subset code.
Can I run it on Windows or macOS?
The README says to use Linux, and the setup guide is written for Linux. The Python package also supports Python 3.13, but the official docs do not document Windows or macOS as supported benchmark hosts.
What license applies?
Harness² is Apache 2.0. Third-party code under third_party/, the benchmarks and their datasets retain upstream licenses. WorkBuddy-Bench uses a custom Tencent license with a territorial limitation, so read it before downloading.
Is it an officially supported Google product?
No. The README explicitly says it is not an officially supported Google product and is not eligible for the Google Open Source Software Vulnerability Rewards Program.
Do I need model credentials for the demo?
No. The README says harness2-demo requires no model credentials. Credentials are needed for model-backed benchmarks, judging and reporting.
How do I resume a run?
Repeat the same command. Results are stored under runs/, and the saved configuration prevents incompatible resumes. Use a new HARNESS2_RUNS_DIR when you change models or task selection.
What does --k change?
--k is the task-level recursion budget. --k 1 runs one Propose–Probe–Compose step; larger values scale the search in parallel or sequential mode. See the tutorial for the exact behavior.
Why must judging and reporting use the same parameters as refinement?
The README requires the same benchmark, executor, subset, --k, --mode and seed, because the judge and report must point at the same run. Changing them changes the measurement.
How do I avoid benchmark contamination?
Executor network access is unrestricted and Codex runs with its sandbox disabled. Block huggingface.co, hf.co, *.hf.co, github.com and raw.githubusercontent.com at the network level before evaluation, and report the restrictions you used.
Is there a web UI?
The README documents a command-line workflow: harness2, harness2-setup, harness2-check, harness2-judge and harness2-report. It does not document a web interface.
Where should I start?
If you only want to see the system run, start with the credential-free demo. If you want to reproduce a benchmark result, read Benchmarks first, especially the web-access restriction.
Sources
Repository: github.com/google-research/harness2 · Paper: arXiv 2610.14224 · Official site: harness2.github.io
This page is part of an unofficial guide. Where the official docs do not specify something, this page says so rather than guessing.