Description
Harness2 is a sophisticated framework designed by Google Research for recursive agent harnessing in an open-world environment. It focuses on optimizing the system surrounding AI models, known as the harness, without modifying the model weights themselves. This approach allows for efficient task execution and continuous improvement through a two-level recursion process.
Read More
The framework operates by refining the execution harness for current tasks using a contrastive Propose–Probe–Compose step, while also writing improvements into a persistent harness for future tasks. This dual-level recursion ensures that the system can adapt and improve over time without the need for supervised fine-tuning or altering model parameters.
Harness2 supports various models and executors, including GLM-5 and Gemini 3.5 Flash, with Vertex AI models being a key component of its architecture. The framework is designed to run on Linux with Python 3.12 and Node.js 20+, and it does not require a GPU for its local demo, which can run on a CPU.
The framework includes eight editable components that can be adjusted to optimize performance, such as system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. These components allow users to tailor the harness to specific tasks and environments, enhancing the overall efficiency and effectiveness of the AI system.
Harness2 is not an officially supported Google product and operates under the Apache 2.0 license. It is designed for users who need a flexible and robust framework for AI model execution and improvement, particularly in professional and legal work environments.
- CategoryAI Agent Builders
- Pricing ModelOpen Source
- DeploymentSelf-hosted
- PlatformsLinux
Harness2's Core Features
Recursive agent harnessing
Two-level recursion process
Model weight preservation
Eight editable components
Supports GLM-5 and Gemini 3.5 Flash
Linux and Python 3.12 compatibility
No GPU requirement for demo
Apache 2.0 license
How to use Harness2?
Install: Set up on Linux with Python 3.12 and Node.js 20+
Configure: Adjust editable components for task optimization
Run: Execute local demo without model credentials
Optimize: Use Propose–Probe–Compose for harness improvement
Harness2's Use Cases
- Professional Work
- Legal Work
- Office Automation
- Web Development
- Code Execution
