Harness²
Agent Frameworks Open source (Apache 2.0)
Documentation: github.com · Checked Oct 19, 2026
About
Harness² is a research project from Google Cloud AI Research that improves an agent by rewriting the scaffolding around it: the system prompt, guardrails, memory, tool guidance, sub-agents, skills, scripts, and plug-ins. The model itself stays frozen while the harness proposes, tests, and combines better scaffolding, with no answer key needed. This is research code, not an officially supported Google product, and the local demo runs with no model credentials.
Details
Open source. You run it yourself.
Auth: Open source (Apache 2.0)
Rate limit: Local demo needs no model credentials. Model-backed runs cost whatever your own API usage costs.
Verified: 2026-10-19
Best for: researchers developer
Setup
- Use Linux with Python 3.12 and uv installed
- Clone the repo and install it with the render extras
- Run the local demo to see both recursion modes with no credentials
- Read the setup guide before trying a model-backed run
- Remember this is research code, so expect rough edges
harness2-demo --output runs/demo --k 2Use cases
- Study how scaffolding changes agent behavior without retraining a model
- Run the no-credential demo to see recursive self-improvement in action
- Experiment with self-improving agent setups in a lab setting