agentos.guide › blog
Google Harness2: The Quick Start, And How It Fits Your Agent OS
Google Harness2, written Harness² or Harness Squared, is a free, open-source research framework from Google Research that improves an AI agent by editing the setup around the model while the model stays the same.
The code is on GitHub at google-research/harness2 under the Apache-2.0 licence, and the paper is arXiv 2610.14224.
An agent tries a task, Harness² proposes and tests edits to its setup on that same task, and then it composes the edits that worked into a better setup.
It saves the lessons in an improvement harness, so each new task starts with better ideas about what to change.
This page is the quick start, plus how Harness² fits an Agent OS workflow.
If you want help turning AI agents and content into Google and AI search traffic, book a free SEO strategy session with my team.
Google Harness2 at a glance
Here are the facts I checked on the repo, the paper and the project page before writing this.
| Fact | What the official sources say |
|---|---|
| What it is | Harness² is a plug-and-play framework for recursive agent harnessing, which improves an agent's setup while the model weights stay fixed. |
| Who made it | The authors are from Google Cloud AI Research and Northwestern University, and the repo sits under google-research. |
| Cost | The code is free and open source under Apache-2.0, and model-backed runs incur API costs. |
| Repo | The code lives at github.com/google-research/harness2. |
| Paper | The paper is arXiv 2610.14224, "Harness^2: Recursive Agent Harnessing for an Open World", submitted on 9 October 2026. |
| Version | The pyproject file lists version 0.1.0. |
| Popularity | The repo had over 16,000 stars and 1,500+ forks by 21 October 2026, and it is still climbing. |
| Best reported gains | One step gave up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB. |
| Status | The README says it is not an officially supported Google product. |
What Harness² does, in one minute
The README defines a harness as the system around a model that enables it to act as an agent, including prompts, tools, memory, skills and orchestration.
Harnessing is the process of optimising that system, and Harness² makes that process recursive.
I explain it in my video with a kitchen, where 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.
Task-level recursion improves the harness for the current task through a step called Propose, Probe, Compose.
An agent executor runs the starting harness, a harness editor proposes invariant and domain edits, candidate probes run them, and a harness composer combines them into the final harness.
Harnessing-level recursion then has an experience reflector update a persistent improvement harness with adopted harnesses and evidence-supported lessons.
It needs no answer key, and I compared that to testing two routes to the gym and keeping the faster one.
🔥 Want the exact Harness² setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
Google Harness2 quick start in seven steps
This is the fastest route from nothing to your first judged run.
| Step | What you do | Command or place |
|---|---|---|
| 1 | Get a Linux machine with Python 3.12 and uv installed. | Linux box or server |
| 2 | Clone the repo and move into the repository root. | git clone https://github.com/google-research/harness2.git |
| 3 | Create the environment and install Harness² with its render extras. | uv pip install -e '.[render]' |
| 4 | Run the free local demo, which needs no model credentials. | harness2-demo --output runs/demo --k 2 |
| 5 | Install Node.js 20+ and the Google Cloud CLI, then add your Vertex AI project to .env. | gcloud auth application-default login |
| 6 | Install one benchmark and run a single task. | harness2-setup lab |
| 7 | Judge and report the run with the same settings. | harness2-judge and harness2-report |
These are the README's exact install commands, 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 2The README says the demo runs both recursion modes with deterministic editors and real Python executions, and it requires no model credentials.
For model-backed runs, these are the README's configuration commands.
cp .env.example .env
# Edit .env with your Vertex AI project and model IDs.
gcloud auth application-default loginThis is the README's first real run, which installs LAB and runs one 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 parallelThe README warns that model-backed runs incur API costs, so start with a small task selection like this.
JobBench installs with harness2-setup jb, and WorkBuddy-Bench installs with harness2-setup wb and takes --subset office, --subset web or --subset code.
LAB needs Pandoc and Poppler, and WorkBuddy-Bench needs Docker.
If you want my team to look at how AI agents and AI search can grow your business, book a free one-to-one GEO, AEO AI audit.
The eight harness parts, mapped onto an Agent OS
The README says Harness² exposes eight editable components around the model.
This is the table I'd pin on the wall of any Agent OS, because every row is something you already have.
| Harness part | Runtime role | What it looks like in my Agent OS |
|---|---|---|
| System prompt | It constructs context. | It is the main instruction file each agent reads first. |
| Guardrails | It constructs context. | It is the list of rules an agent must never break. |
| Memory | It constructs context. | It is the shared Obsidian vault that Claude, Hermes, OpenClaw and OpenCode all read. |
| Skills | It constructs context. | It is the skill files that tell an agent how to do one job. |
| Tool guidance | It defines the tool interface. | It is the notes that explain when and how to use each tool. |
| Scripts | It defines the tool interface. | It is the ready-made scripts an agent can call instead of improvising. |
| Subagents | It orchestrates delegated work. | It is the helper agents a lead agent hands work to. |
| Plugins | It controls middleware. | It is the add-ons that run inside the tool loop. |
In my Agent OS, Claude, Hermes, OpenClaw and OpenCode share one memory in Obsidian.
When something breaks, the fix is usually in the setup rather than the model.
It is usually a skill file with a bad rule, a missing memory note or vague instructions.
OpenCode runs inside my Agent OS with MiniMax M3.1 Flash, and OpenCode is also the default executor in Harness².
🔥 Want the exact Harness² setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
How Harness² fits an Agent OS workflow
Harness² is a research framework, so I wouldn't plug it into your daily agents and walk away.
The smart way to use it is as a lab next to your Agent OS.
First, run the free demo so you understand Propose, Probe, Compose.
Second, run one small paid task with OpenCode and read the adopted harness and the reflection notes in the runs folder.
Third, copy any rule, memory note or skill instruction that clearly helped into your own Agent OS files.
Fourth, keep a simple log of the setup changes you make, because that is the hand-made version of the improvement harness.
The biggest lesson for non-coders is that instructions, rules, memory and skills are mostly plain English, and you can improve them today.
If you want to do that by hand first, my guide to the best Hermes Agent skills is the place to start.
If you want a done-with-you plan for turning agents and content into rankings, grab a free one-to-one strategy session and we'll look at your site, your competitors and the keywords you're closest to winning.
Parallel or sequential recursion
The --k flag sets the task-level recursion budget, and the mode decides how it gets spent.
| Option | What it does |
|---|---|
--k 1 | It runs one Propose, Probe, Compose step. |
--k K --mode parallel | It probes K general and K domain-specific candidates, then produces K compositions. |
--k K --mode sequential | It refines the harness through K successive steps. |
--substrate cx | It uses Codex instead of the default OpenCode executor. |
--dry-run | It previews task selection without any model calls. |
The README says parallel explores wider and sequential refines deeper.
In my video I used a budget of 3 as the example, where parallel tries three ideas side by side and sequential improves one idea three times.
The project page shows Gemini 3.5 Flash on OpenCode going from 72.4 to 77.5 on WorkBuddy with one step, and to 80.8 with parallel recursion at k = 5.
Judging and reporting must use the same benchmark, executor, subset, --k, --mode and seed as the refinement run.
Results are saved under runs/, repeating a command resumes it, and a new HARNESS2_RUNS_DIR keeps runs apart when you change models.
Is this just Hermes redone?
A viewer asked exactly that in the comments, and it is a fair question.
Hermes Agent has skills and memory that you can edit yourself as you work.
Harness² is a research framework that automates harness edits for OpenCode or Codex executors and measures them on benchmarks.
I see them as partners, because Hermes does the daily work and Harness² shows which setup changes are worth making.
My Harness Multiplier guide shows the same idea from another angle, which is the same AI getting better results from a better harness.
Caveats before you start
It costs more to run, because each probe is a full run of the task and the README warns about API costs.
It is a research project, and version 0.1.0 is an early release.
It is technical, because it needs Linux, Python 3.12, Node.js 20+ and the Google Cloud CLI.
The README says executor network access is unrestricted and Codex runs with its sandbox disabled, so configure network restrictions before evaluation.
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.
An engineer called Ash Pachar, who used to work at Lightning AI, made a point I paraphrased in my video, which is that a person usually designs a harness once and leaves it.
Harness² is built to stop that from happening.
Also On Our Network
FAQ
What is Google Harness2?
Google Harness2, written Harness² or Harness Squared, is a free, open-source research framework from Google Research at github.com/google-research/harness2. It improves an AI agent by editing the harness around the model, such as its system prompt, memory and skills, while the model weights stay fixed.
How do I install Harness²?
The README says to use Linux, Python 3.12 and uv. From the repository root you run uv venv --python 3.12 .venv, source .venv/bin/activate, uv pip install -e '.[render]' and harness2-demo --output runs/demo --k 2 for the free local demo.
Is Harness² free?
Yes. The code is free and open source under Apache-2.0, and the local demo needs no model credentials. Model-backed benchmark runs go through Vertex AI, and the README warns that they incur API costs.
Which agents does Harness² work with?
OpenCode is the default executor, and the --substrate cx flag switches to Codex. The paper reports gains for GLM-5 and Gemini 3.5 Flash across four model and harness configurations.
Is Harness² just Hermes Agent redone?
No. Hermes Agent is an everyday agent whose skills and memory you edit yourself. Harness² is a research framework that automates harness edits for OpenCode or Codex executors and measures them on benchmarks.
Is Harness² an official Google product?
No. The README says it is not an officially supported Google product, and the pyproject lists it as version 0.1.0, an early research release.
Final word on Google Harness2
Run the free demo, read one adopted harness, and copy the best rule into your own agent today.
If you want a clear plan for getting your business found in Google and AI search, book your free strategy session here.
If you want the whole Agent OS with shared memory, skills and agents already wired in, it lives inside the AI Profit Boardroom, with 3,400+ members and four coaching calls a week.
That is the fastest way to put the big idea behind Google Harness2 to work.
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert and the founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.
Related guides on agentos.guide
→ The Harness Multiplier + Pi Agent
→ DeepSeek Harness: Everything Is A Plugin
→ The Book Brain Engine For Hermes
→ Hermes AI + Obsidian Second Brain
Get the whole Agent OS
The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.
Join the AI Profit Boardroom →📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉