AI Breakthrough: Self-Improving Branches for Smarter Agents

Key Summary

Researchers from Meta, Duke University, and the University of California, Davis have developed a new method for improving AI agent performance by optimizing the scaffolding around the model, rather than the model itself. The approach, titled "Mixture of Self-Improving Branches for Agent Harness Optimization," achieved a 34.8% relative improvement in Olympiad-level math reasoning, from 46.0% to 62.0%, using the Gemini 3 Flash model without retraining the model. The new method, built on the earlier Meta-Harness system, uses multiple specialized branches that evolve independently, each with its own subset of development data and policy for proposing changes. The results were achieved using development-set data, and the gains carried over to agentic tasks, including a 11.6% gain on Terminal-Bench 2.0 and a 3.8% increase on SWE-bench Lite.