GitHub radar
Automatically improve your AI agent's setup
Harness² is a Google research tool that automatically finds better settings for your AI agent on a given task, without touching the model itself. After a run you get an improved agent configuration and a report of what changed.
It's a system that tries different agent configuration variants on one task and assembles an improved version from the ones that worked. The model stays the same — only the setup around it changes: prompts, memory, tool instructions.
When it helps
Useful if you tune your agent's behavior yourself and want to test changes on real tasks instead of guessing. Skip it if you just hand tasks to an agent and never touch its configuration.
Pros
- Free to try: the local demo runs with no model API keys
- Tests configuration changes on a real task, not just in theory
- Keeps useful settings and reuses them on later tasks
- Open source under Apache 2.0 — change it however you want
Cons
- Real benchmarks need paid Google Cloud Vertex AI access
- Results only cover Gemini and GLM-5, not Opus or Fable
- Useless if you don't tune your agent's configuration yourself
How to set it up — step by step
- 1Open your AI agent (Claude Code, Codex) in the folder you'll work from
- 2Send the agent the text from the block below — it installs the local demo from GitHub
- 3Wait while the agent installs it via uv and runs the harness2-demo command
- 4Check the result in the runs/demo folder — it shows what changes the tool proposed and assembled
- 5For a real model benchmark, arrange paid Google Cloud Vertex AI access first
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the local Harness² demo from https://github.com/google-research/harness2 — follow the README with uv (the harness2-demo command, no model keys needed) — and show me what's in runs/demo. If you want to run a real model-backed benchmark, ask me for Google Cloud / Vertex AI access.
No GPU needed — models are called over the Vertex AI API on Google Cloud, not run locally. The model-free local demo installs on a regular laptop via uv. Running real benchmarks adds Node.js, the Google Cloud CLI and paid Vertex AI access, plus Pandoc, Poppler and Docker for LAB and WorkBuddy-Bench.
Open on GitHub▌ More finds