Google Open-Sources Harness2 for Recursive AI Agents
Google Research has released Harness2, an Apache-2.0 framework that boosts GLM-5 scores on WorkBuddy-Bench by 11.5 points without weight updates.
Stock photo for illustration only, not from the actual event
- Google Research has officially released the Harness2 open-source framework under the Apache-2.0 license.
- The system is designed to rewrite an AI agent's harness dynamically during test time.
- Applying just a single refinement step lifted GLM-5 scores by 11.5 points on WorkBuddy-Bench.
- All performance gains are achieved without requiring any underlying model weight updates.
Google Research has expanded its open-source artificial intelligence ecosystem with the introduction of Harness2, a new framework released under the Apache-2.0 license. This newly unveiled system focuses on reshaping how AI agents manage their operational harness and execution flow specifically during testing phases.
The core architecture of Harness2 introduces a recursive approach to agent harnessing, allowing the system to adapt and optimize its behavior dynamically at test time. By decoupling the harnessing logic from the core neural network training, developers can experiment with advanced agent behaviors without the traditional overhead associated with model retraining cycles.
To demonstrate its practical effectiveness, the research team applied the framework's refinement process to benchmark evaluations. Just a single refinement step successfully elevated the performance of the GLM-5 model by 11.5 points on WorkBuddy-Bench, showcasing the immediate impact of the recursive harnessing methodology.
Stock photo for illustration only, not from the actual event
The ability of Harness2 to enhance agent capability without modifying model weights represents a significant shift in deployment strategies. Traditional scaling often demands costly retraining phases, whereas test-time harness rewriting allows engineers to optimize existing models on the fly, offering a flexible and resource-efficient pathway toward more autonomous and reliable AI workflows.
By open-sourcing this framework, Google aims to empower the broader developer and research community to experiment with recursive agent architectures, potentially accelerating innovation across automated task execution and complex problem-solving domains.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment