Google Open Sources Harness2 for Self Improving AI Agents
Google has open-sourced Harness2, a framework that enables AI agents to iteratively self-improve their operational harnesses—including system prompts, tools, and memory—across tasks without modifying the underlying model weights. This release allows developers to enhance agent performance and adaptability over time without the immense cost of retraining or fine-tuning the underlying model weights. It represents a significant step forward in agentic workflows by optimizing the surrounding execution environment dynamically. Harness2 allows agents to propose changes, test candidate versions, and select better setups by comparing performance without requiring correct answers or graders. Google tested the framework using models like Gemini 3.5 Flash and GLM-5, demonstrating performance gains across legal, office, and professional benchmarks.
## BACKGROUND
An AI agent harness refers to the execution layer and surrounding environment that determines what a model observes, what tools it can invoke, and how instructions or workflows are organized. Traditional agent setups rely on fixed configurations designed manually by humans, which limit the agent's ability to adapt to new tasks.