google-research/harness2

What is it?

Key features

• 01Task-level recursion refines the harness via Propose–Probe–Compose without ground-truth feedback.
• 02Harnessing-level recursion keeps an improvement harness of adopted harnesses and lessons across tasks.
• 03Parallel and sequential recursion modes control how far and how deep refinement goes.
• 04Supports OpenCode and Codex executors.
• 05Includes JobBench, LAB and WorkBuddy-Bench benchmark integrations.
• 06Edits eight harness components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.

Use cases

• Researchers reproducing harness optimization experiments on JobBench, LAB or WorkBuddy-Bench.
• Teams comparing agent harness designs for a fixed model across professional task benchmarks.
• Developers studying how prompts, tools and orchestration affect agent performance without retraining the model.

Quick start

1. 1

   Create a Python 3.12 virtual environment

   uv venv --python 3.12 .venv
2. 2

   Activate the virtual environment

   source .venv/bin/activate
3. 3

   Install the package in editable mode with render extras

   uv pip install -e '.[render]'
4. 4

   Run the local demo (no model credentials needed)

   harness2-demo --output runs/demo --k 2

Things to know

• !Model-backed benchmarks need Vertex AI setup, Google Cloud CLI authentication, and possibly Docker, Pandoc and Poppler.
• !Model-backed runs incur API costs.
• !Executor network access is unrestricted, so benchmark answers may need network restrictions configured beforehand.

Topics

FAQ

What is harness2?

A Python framework for recursive harness optimization that refines the prompts, tools and orchestration around a fixed LLM agent on unseen tasks. Harness² is a framework for improving the system around a language model that lets it act as an agent. That surrounding system, called a harness, includes prompts, tools, memory, skills and orchestration. The framework keeps model weights fixed and instead optimizes the harness, aiming at tasks the agent did not encounter during harness design.

Is harness2 free to use? What license does it have?

harness2 is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.

Is harness2 still maintained?

Actively maintained. The last commit to harness2 was on Oct 14, 2026.

What can I use harness2 for?

Researchers reproducing harness optimization experiments on JobBench, LAB or WorkBuddy-Bench. Teams comparing agent harness designs for a fixed model across professional task benchmarks. Developers studying how prompts, tools and orchestration affect agent performance without retraining the model.

How do I install harness2?

Create a Python 3.12 virtual environment: uv venv --python 3.12 .venv Activate the virtual environment: source .venv/bin/activate Install the package in editable mode with render extras: uv pip install -e '.[render]' Run the local demo (no model credentials needed): harness2-demo --output runs/demo --k 2

Where can I get harness2?

The source code and installation instructions are on GitHub: https://github.com/google-research/harness2