Google Open-Sources Harness² Framework to Enhance AI Agent Performance
Google Cloud AI Research has released Harness², an open-source framework that improves agent performance by optimizing components like prompts and tools without modifying the underlying AI model weights.
Google Cloud AI Research has introduced Harness², a new open-source framework designed to improve the operational systems of AI agents. Released under the Apache-2.0 license, this tool allows for the optimization of system components—such as prompts, memory, tools, and skills—during deployment, all without altering the base model's weights.
Harness² manages eight editable components that constitute an AI agent's environment: system prompts, safety boundaries, memory, tool guidance, sub-agents, skills, scripts, and plugins. The system operates through a two-stage structure: task-level recursion using a 'Propose, Test, Create' method, and harness-level recursion, which stores experiences in a file to guide future tasks.
Tested on various model setups such as Gemini 3.5 Flash and GLM-5, the framework demonstrated performance improvements in benchmarks including WorkBuddy-Bench, JobBench, and Harvey’s LAB. With this plug-and-play tool, developers can enhance the capabilities of their existing agents without needing a more powerful model.