Google Harness2 — written Harness², or “harness squared” — is a free, open-source framework from Google Research that makes AI agents better without a new model. Instead of retraining anything, it rewrites the setup around the agent: the prompt, the guardrails, the memory, the tools and the skills. It then tests the changes and keeps what works. In the video below I break down how it works and why it matters for anyone running agents. Here’s the plain-English version, plus how to use the idea in your business today.
Key takeaways
- Google Harness2 improves the harness around an AI agent, not the model.
- It runs a propose → probe → compose loop with no answer key needed.
- Reported gains of up to +11.5 points on WorkBuddy-Bench from one step.
- Free and open source (Apache 2.0) — claim a free strategy session to apply it to your business.
What is Google Harness2?
Most people think a better AI agent needs a better model. Google Harness2 takes the opposite view. The model stays frozen; the harness — everything wrapped around the model — is what gets better. That harness is the instructions, the rules, the memory, the tool guidance and the skills that decide how the agent behaves.
Harness² was built by researchers at Google Cloud AI Research with Northwestern University and published with the paper Recursive Agent Harnessing for an Open World. The code is on GitHub under the Apache 2.0 licence, and it picked up around 15,000 stars within days of launch.
| Google Harness2 at a glance | |
|---|---|
| Full name | Harness²: Recursive Agent Harnessing for an Open World |
| From | Google Cloud AI Research, with Northwestern University |
| What it does | Improves the harness around an AI agent while the model stays frozen |
| Core loop | Propose → Probe → Compose |
| Editable parts | System prompt, guardrails, memory, tool guidance, sub-agents, skills, scripts, plugins |
| Executors | OpenCode (default) and Codex |
| Models tested | Gemini 3.5 Flash and GLM-5 |
| Licence | Apache 2.0, free and open source |
| Status | Research code — “not an officially supported Google product” |
| Links | GitHub · project page · paper |
How Google Harness2 works: propose, probe, compose
Each improvement round follows three steps:
- Propose — a harness editor suggests changes. Some are general (better tool guidance), some are specific to the task in front of it.
- Probe — each candidate harness is actually run on the task, so the agent tries it for real.
- Compose — a harness composer compares the runs and merges what worked into one improved successor harness.
The clever part is that it doesn’t need ground-truth answers. It compares runs against each other — what the paper calls a contrastive step — so it can improve on open-ended work where there is no answer key. That’s exactly the kind of messy work most of us give agents.
Want agents that get better every week, not worse? Claim your free strategy session →
Want self-improving agents working in your business? Inside the AI Profit Boardroom you get my Agent OS — Claude, Hermes, OpenClaw and OpenCode sharing one memory and one set of skill files — plus four live coaching calls a week.
Want a plan built around your business first? Claim your free strategy session.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →
The 8 parts Google Harness2 can edit
Harness² can change eight components of an agent’s setup:
- System prompt — the core instructions.
- Guardrails — what the agent must and mustn’t do.
- Memory — what it remembers and reuses.
- Tool guidance — when and how to use each tool.
- Sub-agents — helpers it can hand work to.
- Skills — reusable how-to files.
- Scripts — code it can run.
- Plugins — extra abilities it can load.
If you run Claude Code, Hermes or OpenCode, you’ll recognise every one of these. They’re the same levers you already pull by hand — Harness² just pulls them automatically and keeps score.
Two loops: why it’s called harness squared
The “squared” comes from two levels of recursion. Task-level recursion improves the harness for the job in hand, either wider (several candidates in parallel) or deeper (one candidate after another). Harnessing-level recursion goes a step further: an experience reflector saves the winning setups and the lessons backed by evidence into an improvement library, so the system gets better at improving itself on future tasks.
Each new task still starts from the base harness, but the improver behind it has learned from every task before. That’s the self-improving part.
Google Harness2 results
Google tested Harness² with two models — Gemini 3.5 Flash and GLM-5 — across four model-and-harness combinations, using OpenCode and Codex as the agent executors. A single improvement step consistently helped, and more rounds helped more.
| Benchmark | Best reported gain over the base harness |
|---|---|
| WorkBuddy-Bench (code, web, office) | Up to +11.5 points |
| JobBench | Up to +4.5 points |
| LAB | Up to +11.8 percentage points |
| WorkBuddy, Gemini 3.5 Flash + OpenCode | 72.4 → 77.5 after one step; 80.8 with five parallel probes |
These are the researchers’ own reported numbers on their benchmarks. They’re a strong signal, not a promise for your tasks — but they show how much performance is sitting in the harness rather than the model.
Want to know where agents could save your team hours? Claim your free strategy session →
How to try Google Harness2
The repo is research code, so it’s built for technical users. According to the README you need:
- Linux and Python 3.12, set up with uv.
- Node.js 20+ for model-backed benchmarks.
- Google Cloud with Vertex AI and the gcloud CLI for models.
- Docker for WorkBuddy-Bench, and Pandoc and Poppler for LAB.
There’s a demo that runs without any model credentials, which is the safe place to start. OpenCode is the default executor; Codex is an option.
Watch your costs. Every probe is a full agent run, so model-backed runs add up fast — the README says to start with a small task selection. Executor network access is unrestricted by default, so lock it down before running evaluations, and remember this is research code, not a supported Google product.
What Google Harness2 means for your business
You don’t need to run Harness² to steal its biggest lesson: the setup around your agent matters as much as the model. Most businesses switch models when an agent disappoints. The faster win is usually a better prompt, clearer tool guidance, tighter guardrails and good skill files — then testing changes side by side and keeping the winners.
That’s how I run my own Agent OS. Claude, Hermes, OpenClaw and OpenCode share one memory and one set of skill files, so when I improve a skill, every agent gets better at once. Harness² is Google’s research-grade version of the same idea, with the testing automated.
- Write your rules down as skill files instead of retyping prompts.
- Change one thing at a time and compare the results.
- Keep a lessons file of what worked, so every agent learns from it.
- Give agents only the access they need.
My verdict on Google Harness2
Google Harness2 is one of the most important agent releases of the year, because it proves you can get real gains without waiting for a new model. If you’re technical, try the demo. If you’re a business owner, apply the lesson now: improve the harness, test, keep what works. Want help doing that? Claim your free strategy session.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Claude, Hermes, OpenClaw and OpenCode together in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
FAQ: google harness2
What is Google Harness2?
Harness² is a free, open-source framework from Google Research that improves an AI agent’s harness — prompt, guardrails, memory, tools, skills and more — while the model stays frozen.
Does Google Harness2 train a new model?
No. The model weights don’t change; only the setup around the model is edited and tested.
How much does Google Harness2 improve agents?
The paper reports gains of up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB from one improvement step.
Is Google Harness2 free?
Yes — it’s open source under Apache 2.0, but model-backed runs cost API credits because every probe is a full agent run.
Which agents does Google Harness2 work with?
OpenCode is the default executor and Codex is supported; Google tested it with Gemini 3.5 Flash and GLM-5.
Is Google Harness2 an official Google product?
No — the repo says it is not an officially supported Google product. It’s research code at an early stage.
Ready to put better harnesses around your AI agents? Claim your free strategy session and we’ll map out where agents can save you time and win you traffic.
Want the full system? Join the AI Profit Boardroom for the Agent OS and daily updates.
Getting your site cited by AI too? Get the free GEO & AI SEO checklist.
Claim Your Free SEO Strategy Session →Join the AI Profit Boardroom →Get the Free GEO & AI SEO Checklist →
Related reading
Last updated October 2026. This page is a living guide to google harness2 — the facts here move fast and it is updated as they do.