Open-source framework for recursive agent harnessing: agents that improve their own execution.
Harness² project site
1 / 2Overview
What is Harness²?
Harness the agent. Then harness its self-improvement.
Harness² is a recursive agent harnessing framework for an open world, from Google Cloud AI Research and Northwestern University.
Harness the agent, then harness its self-improvement. Harness² introduces two levels of recursion: task-level recursion improves the execution harness, and harnessing-level recursion guides how future harnesses are improved. The project site walks through the method, recursion, components, results and citation, and the framework is evaluated on WorkBuddy-Bench across Code, Web and Office domains.
The framework is open source under Apache-2.0, implemented in Python, and published on GitHub. It is aimed at AI researchers and agent developers working on self-improving agents.
Platforms and languages
Harness² Availability
Platforms
Languages
Capabilities
Harness² Key Features
Recursive agent harnessing
Task-level recursion improves the execution harness while harnessing-level recursion guides how future harnesses are improved.
Open world agent evaluation
Benchmarks agents across Code, Web and Office domains with WorkBuddy-Bench.
Self-improvement at scale
Task-level scaling from k=1 to 5 shows how harness improvements compound.
Open source from Google Research
Apache-2.0 licensed Python framework from Google Cloud AI Research and Northwestern University.
Best for
Who uses Harness²?
AI researchers and agent developers
Build and evaluate self-improving agent harnesses with a recursive framework.
Plans and access
Harness² Pricing
Free
Common questions
Harness² FAQs
What is Harness²?
An open-source framework for recursive agent harnessing: harness the agent, then harness its self-improvement.
Who built it?
Researchers at Google Cloud AI Research and Northwestern University.
Is it free?
Yes. It is open source under the Apache-2.0 license.
How is it evaluated?
On WorkBuddy-Bench across Code, Web and Office domains, with task-level scaling from k=1 to 5.
Where can I read more?
The project site at harness2.github.io covers the method, recursion, components, results and citation.
Reviews
See what the community thinks and share your experience.
—
Based on 0 ratings
Rating distribution
Community reviews
No reviews
No written reviews yet.
