Researchers from Meta, Duke University, and the University of California, Davis have introduced a new method for optimizing AI agent frameworks. The approach, detailed in a preprint titled 'Mixture of Self-Improving Branches for Agent Harness Optimization,' focuses on improving the scaffolding around large language models rather than retraining the models themselves.
The system replaces a single search path with multiple specialized branches that evolve independently using distinct subsets of development data. A router mechanism selects the most appropriate branch for each incoming task. This process is conducted entirely on development-set data without access to the test set.
The new method builds upon the Meta-Harness system, which was released on March 30, 2026. In tests using the Gemini 3 Flash model, the researchers achieved a 34.8% relative improvement in Olympiad-level mathematical reasoning, with accuracy increasing from 46.0% to 62.0%.
Performance gains were also observed in other areas, including an 11.6% improvement on Terminal-Bench 2.0, which evaluates agents in command-line environments. The paper, published on September 29, 2026, as arXiv:2609.37834v1, lists Haoyu Dong, Zihao Lin, Lizhu Zhang, and Zhuokai Zhao as authors.
