Google Harness2, written Harness² or Harness Squared, is a free, open-source framework from Google Research that makes an AI agent better at real work by improving the setup around the model, not the model itself.
It lives on GitHub at google-research/harness2, it uses the Apache-2.0 licence, and the code is free to download.
Here's how it works in three lines.
An agent attempts a task with its current setup.
Harness² proposes changes to that setup, tests them on the same task, and combines the changes that helped into a better setup.
It then saves the lessons in an "improvement harness", so the next task starts with smarter ideas about what to change.
This post is written for business owners, agency owners and creators, so I'll explain what Harness² is first and then show you seven real ways a business can use it or its lessons.
If you want my team to look at how AI agents and AI search can bring your business more customers, book a free one-to-one GEO, AEO AI audit.
Google Harness2 In One Table
Here is the whole thing on one screen, checked against the official repo, paper and project page.
| Question | Answer |
|---|---|
| What is it? | Harness² is a framework for recursive agent harnessing, which means improving an agent's setup again and again without changing the model. |
| Who made it? | Researchers from Google Cloud AI Research and Northwestern University made it, and the repo sits under google-research on GitHub. |
| What does it cost? | The code is free and open source under Apache-2.0, but real runs call paid models. |
| What does it change? | It edits the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. |
| Which agents does it drive? | It drives OpenCode by default, and Codex is an option. |
| What did it achieve? | One step gave up to +11.5 points on WorkBuddy-Bench, +4.5 on JobBench and +11.8 percentage points on LAB. |
| Is it official Google software? | The README says it is not an officially supported Google product. |
| How popular is it? | The repo had over 16,000 stars and 1,500+ forks by 21 October 2026, and it is still climbing. |
🔥 Want my exact How I'd Use Google Harness2 In A Business setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
What Google Harness2 Is, Explained Like You Run A Business
Every AI agent has two parts, which are the model and the harness.
The README describes the harness as the system around the model that lets it act as an agent, including prompts, tools, memory, skills and orchestration.
I explain it in my video with a restaurant.
The model is the chef, and the harness is the kitchen, with recipes on the wall, rules, memory, tools, skills and helpers.
A great chef in a messy kitchen still serves bad food.
When an agent gives poor results, most businesses go and buy a more expensive chef.
Harness² tidies the kitchen instead, and it does it automatically.
The Problem It Solves For A Business
Here's the before state most teams are stuck in.
Somebody sets up an agent once, writes the instructions once, and then nobody touches the setup again.
An engineer called Ash Pachar, who used to work at Lightning AI, made a point that I paraphrased in my video, which is that a person usually designs the harness once and then it just sits there.
The agent never gets better at your work, even after hundreds of tasks.
The obvious fix is a bigger model, which costs more every single month.
Harness² tests a different idea, which is that the setup can keep improving while the model stays the same.
The paper backs that up, because the model weights stay fixed and the scores still go up.
How It Works In Plain English
The README describes two loops.
The first loop works inside one task.
An agent executor tries the task with the starting setup, then a harness editor proposes two kinds of changes, which are general "invariant" changes and changes specific to that type of work.
Candidate probes run those changed setups on the same task, and a harness composer combines what worked into a final setup.
This is called Propose, Probe, Compose.
The second loop works across tasks.
An experience reflector looks at what happened and updates the improvement harness with the adopted setups and "evidence-supported lessons", in the README's words.
The key trick is that it doesn't need an answer key.
I compared it to working out the best route to the gym, because you can try two routes and see which is faster without anyone telling you the right answer first.
How To Get Started With Harness²
If you or someone on your team is technical, these are the README's exact install steps for Linux with Python 3.12 and uv, run from the repository root.
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2
The demo runs both recursion modes with deterministic editors and real Python executions, and it needs no model credentials.
That means your developer can show you the whole loop before you spend anything on API calls.
For real runs, the README asks for Node.js 20+, the Google Cloud CLI and a .env file with your Vertex AI project and model IDs.
cp .env.example .env
# Edit .env with your Vertex AI project and model IDs.
gcloud auth application-default login
The README's first real example installs LAB and runs a single task.
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 model-backed runs incur API costs, so start with a small task selection exactly like that.
🔥 Want an agent setup that improves without a developer? Inside the AI Profit Boardroom, I've got the installable Agent OS plus step-by-step tutorials on writing better agent instructions, rules, memory and skills. You'll be learning alongside 3,400+ members with four coaching calls every week. → Get access here
🚀 Ready to put How I'd Use Google Harness2 In A Business to work in your business? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
7 Ways A Business Can Use Google Harness2
None of these are income claims, and they're simply the jobs where Harness² or its lessons fit a real business.
1. Audit Your Agent Setup With The Eight-Part Checklist
The README lists eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.
That list is a free audit checklist for any agent your business already runs.
Go through each part and ask whether it is clear, current and written down.
You don't need to install anything to do this, and it's the fastest win in this whole post.
2. Test A Setup Fix Before You Pay For A Bigger Model
Before you upgrade every agent to a pricier model, ask whether the setup is the real problem.
The project page shows Gemini 3.5 Flash on OpenCode rising from 72.4 to 77.5 on WorkBuddy with one Harness² step, and to 80.8 with parallel recursion at k = 5.
That is the same model doing better work because the kitchen got cleaner.
3. Improve Agents That Do Office Work
WorkBuddy-Bench covers office, web and code tasks, and JobBench covers professional work.
If your team uses OpenCode or Codex agents for spreadsheets, documents or research, those are the closest benchmarks to your daily jobs.
A technical team member can run one small task, read the adopted harness in the runs folder and copy the useful rules into your own setup.
4. Learn From Legal And Professional Workflows
LAB covers legal work, and the paper reports a gain of up to 11.8 percentage points on its partial pass rate.
If you run a professional services business, the adopted harnesses and reflection notes show you what kind of instructions helped an agent on careful, rule-heavy work.
Treat that as research to learn from, not as a ready-made legal assistant.
5. Prove Your Agent Work To Clients
If you build agents for clients, Harness² gives you a before-and-after method.
You run the base setup, you run the improved setup, and you compare the judged results with the same benchmark, settings and seed.
The README insists that judging and reporting use the same benchmark, executor, subset, --k, --mode and seed as the refinement run, which keeps your comparison honest.
6. Fix The Setup In Your Own Agent OS
This is the one I actually live with every day.
In my Agent OS, Claude, Hermes, OpenClaw and OpenCode share one memory in Obsidian.
When something goes wrong, the fix is usually in the setup, such as a skill file with a bad rule, a missing memory note or vague instructions, and not a new model.
OpenCode runs inside my Agent OS with MiniMax M3.1 Flash as its brain, and that is the same executor Harness² uses by default.
If you want to see how that shared memory works, this video walks through Hermes with Obsidian.
7. Train Your Team To Think Like Harness²
The biggest lesson for non-coders is that the setup around the model matters.
Instructions, rules, memory and skills are mostly plain English, and anyone on your team can improve them today.
Give your team the chef and kitchen picture, then ask them to fix one kitchen problem a week.
If you want help working out which agent jobs will actually move your revenue and your search traffic, grab a free strategy session with my team and we'll look at your site, your competitors and the keywords you're closest to winning.
Parallel Or Sequential: Which Should A Business Pick?
Harness² has two ways to spend its budget, which is set with the --k flag.
| Mode | What it does | When a business would pick it |
|---|---|---|
| Parallel | It probes K general and K domain-specific candidates, then produces K compositions. | You'd pick it when you want to explore lots of different ideas at once. |
| Sequential | It refines the harness through K successive steps. | You'd pick it when you have one decent setup and want to polish it. |
| Codex executor | The --substrate cx flag swaps OpenCode for Codex. |
You'd pick it when your team already works in Codex. |
The README says parallel explores wider and sequential refines deeper.
In my video I used a budget of 3 as the example, so parallel tries three ideas side by side while sequential improves one idea three times.
Costs, Limits And Honest Warnings
It costs more than a normal agent run, because every probe is a full run of the task and the README warns about API costs.
It is a research project, and the pyproject lists version 0.1.0, so expect rough edges.
It is technical, because it needs Linux, Python 3.12, Node.js 20+ and the Google Cloud CLI, and LAB also needs Pandoc and Poppler while WorkBuddy-Bench needs Docker.
The README says executor network access is unrestricted and that Codex runs with its sandbox disabled, so never point it at sensitive client data without locking that down.
The README tells you to configure network restrictions before evaluation, because public benchmark rubrics may be accessible online.
The README says it is not an officially supported Google product.
I didn't run Harness² myself in my video, so the results here come from the paper, the README and the project page.
If you would rather have my team map out how your business can win in Google and AI search, book a free SEO strategy session and we will build the plan with you.
🔥 Want to turn better agent setups into real business results? The AI Profit Boardroom gives you the Agent OS, the skills and memory training, and four weekly coaching calls where you can bring your agent setup and get a straight answer. → Join the AI Profit Boardroom
Who Should Use Google Harness2
Agencies and freelancers who build agents for clients should get a developer to run the demo and one small task.
Teams already using OpenCode or Codex for office, research or legal-style work should look at the adopted harnesses it produces.
Solo business owners who don't code should use the eight-part checklist today and skip the install for now.
If you want ideas for where agents fit your business, my best Hermes Agent use cases post lists the jobs I'd start with.
If your agent keeps forgetting things, my best Hermes Agent memory guide fixes the part of the harness most people ignore.
Related Reading
- My best Hermes Agent use cases post shows the business jobs I'd hand to an agent first.
- My best Hermes Agent memory guide explains how to give your agent a memory that lasts.
- My best Hermes Agent skills guide covers the skills worth adding to your setup.
Also On Our Network
- 🌐 The five-step Harness² setup walkthrough
- 🌐 My feature-by-feature Harness² review with scores
- 🌐 How Harness² fits an Agent OS workflow
- 🌐 The models behind the Harness² paper, checked on GoldieBench
FAQ: Google Harness2 For Business
What is Google Harness2 in simple terms?
Google Harness2, or Harness², is a free research framework from Google Research that improves an AI agent by rewriting the setup around the model.
That setup includes its instructions, rules, memory and skills, and the model itself is not retrained.
Can a non-technical business owner use Harness²?
Not directly, because it needs Linux, Python 3.12, Node.js 20+ and the Google Cloud CLI.
You can still use its main lesson today by improving your agent's instructions, rules, memory and skills by hand.
How much does Harness² cost to run?
The code is free under Apache-2.0, and the local demo needs no model credentials.
Real benchmark runs call models through Vertex AI, and the README warns that they incur API costs, so start with one small task.
What results did Harness² get?
In the paper, a single step improved GLM-5 and Gemini 3.5 Flash over their base harnesses.
The gains reached up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB's partial pass rate.
Did Julian test Harness² himself?
No, I explained Harness² in my video from the paper, the README and the project page, and I didn't run it on camera.
My own experience is with my Agent OS, where most fixes come from improving the setup rather than switching models.
If you'd like my team to look at how AI agents and content can win your business Google and AI search traffic, book your free SEO strategy session here.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
About Julian
I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
- I've built a 7-figure SEO and link-building agency, Goldie Agency, with 70+ people in Bangkok.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote the book "Link Building Mastery", which covers nine link building strategies.
- I run GoldieBench, where I score AI models on real one-shot builds.
→ Get my best AI training inside the AI Profit Boardroom
Fix one part of your agent's kitchen this week, and you'll already be using the biggest idea behind Google Harness2.
