Meta, Duke, California University unveil self-improving agent harness optimization

Meta: Meta is a major technology company conducting extensive AI research through its dedicated labs and collaborations with academic institutions. In this news, Meta researchers contributed to a new arXiv paper proposing an improved method for agent harness optimization that builds directly on the company's prior Meta-Harness work. The paper demonstrates how branching the self-improvement process can overcome limitations in single-trajectory evolution. Meta-Harness: Meta-Harness is Meta's iterative method for self-improving AI agent harnesses through repeated code generation and evaluation on a fixed development set. The new paper identifies its single-path search limitation and introduces an adaptive multi-branch alternative with evolving objectives. Experiments show the branched approach yields higher performance across math and coding benchmarks when combined with a router. Gemini 3 Flash: Gemini 3 Flash is Google's lightweight and efficient large language model designed for fast reasoning and agentic tasks. The research paper used this model to benchmark the new harness optimization system against Meta-Harness on Olympiad-level mathematical reasoning. Results highlighted improved performance when routing between specialized harness branches. Duke University: Duke University is a leading private research university with strong programs in computer science and artificial intelligence. Its researchers collaborated on the paper exploring mixture of self-improving branches for optimizing LLM agent harnesses. The involvement reflects Duke's role in advancing recursive self-improvement techniques through academic-industry partnerships. California University: California University refers to institutions within the University of California system, renowned for cutting-edge work in AI and machine learning. Researchers from the university contributed to developing the branched harness optimization approach described in the paper. Their participation supports the creation of adaptive search methods that produce complementary agent harnesses. Benchmark Scope: The approach was evaluated on mathematical reasoning and agentic coding tasks, showing gains over the prior single-trajectory baseline across multiple settings. Deployment Strategy: A router selects the most suitable branch head for each new input using only development data, allowing complementary harness strengths to be combined without test-set access. Research Innovation: The new method makes the harness improvement process itself adaptive by splitting search into branches that maintain distinct development cases and proposal policies based on their own histories.