What’s new here

Harness² treats the harness as a system to optimize. It works on two levels: it refines the harness for the current task, and it keeps a persistent improvement harness that carries lessons into later tasks. The model weights never change.

What it does

A harness here has eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. For each task, an editor proposes changes, candidate probes run them, and a composer combines them into a final harness. This is the Propose–Probe–Compose step, and it uses contrasts between executions rather than ground-truth feedback.

You scale it two ways. Parallel mode probes K general and K domain-specific candidates and produces K compositions. Sequential mode carries the refined harness through K successive steps. After each task, the adopted harness and the lessons that the evidence supports go into the improvement harness.

The release supports OpenCode and Codex executors and three benchmarks: JobBench, LAB and WorkBuddy-Bench. The README reports that a single Harness² step improves GLM-5 and Gemini 3.5 Flash over their base harnesses, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB’s partial pass rate.

Who it’s for

Researchers and agent builders who want to test automated harness editing against base harnesses on professional-work benchmarks. It is a research release tied to a paper, so expect to work with the benchmark setups.

Try it

You need Linux, Python 3.12 and uv. The local demo needs no model credentials and runs both recursion modes with deterministic editors.

uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2

For model-backed benchmarks you also need Node.js 20+, the Google Cloud CLI and the benchmark dependencies. LAB needs Pandoc and Poppler, and WorkBuddy-Bench needs Docker. Configure your Vertex AI project and model IDs first:

cp .env.example .env
# Edit .env with your Vertex AI project and model IDs.
gcloud auth application-default login

Then run one LAB task, judge it and report. Model-backed runs cost API money, so start small.

harness2-setup lab
harness2-check --bench lab
harness2 --bench lab --k 1 --mode parallel \
  --only-domains antitrust-competition --max-tasks 1
harness2-judge --bench lab --k 1 --mode parallel
harness2-report --bench lab --k 1 --mode parallel

The README warns that executor network access is unrestricted and Codex runs with its sandbox disabled. Configure network restrictions before evaluating, so the agent cannot reach benchmark answers.

How mature is it

The repo was created on 2026-10-06 and last pushed on 2026-10-14. It has 20,342 stars and 1,993 forks, which is a lot of attention for something so new. It has no releases, 2 contributors and 3 commits in the last 90 days. There are 28 open issues and 103 open pull requests. The license is Apache-2.0. The README says it is not an officially supported Google product.