Google releases open-source recursive agent framework to enhance model performance

Google Cloud AI Research has released Harness², an open-source framework designed to improve agent performance without requiring a better model. The framework enables developers to refine an agent’s harness during deployment, focusing on components like prompts, tools, memory, and orchestration. Unlike traditional methods that rely on labels or test cases, Harness² operates in an open-world setting where tasks are unpredictable and no verifier confirms correctness. The framework features eight editable components, extending previous work by adding one more element. It supports two recursive modes—parallel and sequential—that scale search and refinement processes. Testing on benchmarks like WorkBuddy-Bench showed improved results for weaker executors, with GLM-5 outperforming Gemini 3.5 Flash after one step. The project is available on GitHub, with instructions for local and model-backed runs. The framework’s documentation targets Linux, Python 3.12, and uv, with optional dependencies like Node.js, the Google Cloud CLI, and Docker. Users can run both recursion modes without model credentials, though advanced setups require additional tools. The release includes links to the paper, GitHub repository, and project page, with credit given to the researchers. Google encourages engagement through social media and newsletters, expanding its community reach.