Meta, Duke, and UC Davis Researchers Unveil Self-Improving AI Agent Harness Optimization
Researchers from Meta, Duke University, and the University of California, Davis have introduced a new approach to enhance AI agents without altering their core models. Their preprint, titled 'Mixture of Self-Improving Branches for Agent Harness Optimization,' details a system that splits the 'harness search' into multiple specialized branches. An agent harness is the code framework surrounding a large language model, encompassing prompts, tools, context, and action execution. Instead of a single search path, this new method allows each branch to evolve independently, using a distinct subset of development data and its own policy for proposing changes. A router then selects the most suitable branch for each incoming task during deployment. This system demonstrated a 34.8% relative improvement in Olympiad-level math reasoning, increasing accuracy from 46.0% to 62.0% with the Gemini 3 Flash model, without any retraining of the model itself. Significant gains were also observed in agentic tasks, with an 11.6% ...
