HARNESS: Google's Free Tool That Lets AI Agents Upgrade Their Own Setup
Google Research open sourced a tool that improves the prompts, tools, memory and skills around an AI agent, with no retraining. What it is, an honest setup, the catch, and how to use the idea with your own agents
What it is
A harness is everything around the AI model that turns it into an agent: the prompts, the tools, the memory, the skills and the rules. Claude has one too. Most people tune theirs by hand: change a prompt, try it, change it again.
- Google Research released an open-source tool that does that tuning for you. The model itself is never retrained. Only the setup around it changes.
- It works in three moves: it proposes edits to the harness, tests each edit on the same task, then combines the edits that worked into a better harness.
- It also keeps a running file of lessons from every task, so the next round of tuning starts smarter. In other words, it gets better at getting better.
- In Google's own tests, one round raised scores by up to 11.5 points on a work benchmark.
This guide covers an honest setup, the catch (please read it, this is research code), and how to use the same idea with your own agents today.
What you need
- A Linux computer (the README says Linux, Python 3.12 and uv). On a Mac, use a Linux machine or a cloud box.
- Claude Code, if you want Claude to do the install for you.
- For the free demo: nothing else. It needs no model account.
- For real runs: Node.js 20 or newer, the Google Cloud CLI, and a Google Cloud project with Vertex AI. Real runs cost API money.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (5 more sections, 3 copy-paste prompts) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
50,000+ follow where these guides come from. No spam, unsubscribe in one click.