harness2 vs other agent harnesses: what it is and how to start
Harness² (harness2) is an open-source research framework from Google Research that tunes the harness around an agent while the model stays fixed. It runs on top of existing agent harnesses rather than replacing them. This page compares it with OpenCode and Codex CLI using only each project’s own documentation.
What harness2 is
The repository describes Harness² as “Recursive Agent Harnessing for an Open World”: a plug-and-play framework for recursive agent harnessing in open-world settings, with model weights held fixed. It is released under Apache 2.0. The README says it is not an officially supported Google product, and links the paper as arXiv:2610.14224.
How it differs from other harnesses
OpenCode and Codex CLI are coding agents you use directly. Harness² is a layer above them: its README names OpenCode (the default) and Codex as the executors it runs on. So the question is not which one to pick, but whether you want to run an automated harness-improvement loop on top of an agent you already use.
Reported results, not independent tests
The README reports that a single Harness² step, with GLM-5 and Gemini 3.5 Flash over their base harnesses, improved WorkBuddy-Bench by up to 11.5 points, JobBench by up to 4.5 points and LAB by up to 11.8 percentage points of partial pass rate. These are the authors’ own figures; AgentSkillsDir has not run them.
When to use it, and when not to
It suits researchers and teams who already run OpenCode or Codex and want to test automated harness tuning on JobBench, LAB or WorkBuddy-Bench. It is not a ready-made daily coding agent. It needs Linux, Python 3.12 and uv; the model-backed benchmarks also need Node.js 20+ and the Google Cloud CLI, LAB needs Pandoc and Poppler, and WorkBuddy-Bench needs Docker. The local demo needs no model credentials.
Harness² compared with OpenCode and Codex CLI
| Project | Role (from its README) | License | Needs | Official source |
|---|---|---|---|---|
| Harness² | Framework for recursive agent harnessing; model weights fixed; runs on OpenCode (default) or Codex | Apache-2.0 | Linux, Python 3.12, uv; local demo needs no model credentials | google-research/harness2 |
| OpenCode | “The open source AI coding agent” | MIT (per GitHub) | See its README for install options | anomalyco/opencode |
| Codex CLI | “A coding agent from OpenAI that runs locally on your computer” | Apache-2.0 | Install via npm, Homebrew or the install script | openai/codex |
Gemini, Codex and OpenCode are named only to identify those projects. Anything not stated in an official source is left out.
Install and get started (official sources)
Harness²
Run from the repository root. This is the README’s local demo and needs no model credentials.
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2Commands are copied from the official READMEs. Read them before running anything; nothing is installed from this page.
FAQ
What is harness2?
Harness² is an open-source framework from Google Research for recursive agent harnessing: it improves the harness around an agent while the model weights stay fixed. It is released under Apache 2.0 and is not an officially supported Google product.
Is harness2 a replacement for OpenCode or Codex?
No. Its README names OpenCode (default) and Codex as the executors it runs on, so it sits on top of them.
How do I install harness2?
From the repository root, create a Python 3.12 environment with uv, install it with uv pip install -e '.[render]' and run harness2-demo. The exact commands are in the install section of this page and in the official README.
Are the benchmark gains independently verified?
No. The figures come from the project’s README; AgentSkillsDir has not run the benchmarks.
Reference: https://github.com/google-research/harness2
These selections are organized by task, without invented ratings, search volumes or compatibility-test results.