Mixture of Self-Improving Branches For Agent Harness Optimization

2026-09-29

AI Summary

This study addresses the limitation of fixed strategies in agent framework optimization, which are prone to converging on local optima. To overcome this, we propose a recursive self-improvement method based on a multi-branch architecture. This approach integrates adaptive branch search, a dynamically evolving development set, and proposal strategies, employing an input routing mechanism to enable the intelligent selection and co-evolution of complementary frameworks. Experimental results demonstrate that the proposed framework significantly outperforms the Meta-Harness baseline on mathematical reasoning and code generation benchmarks, achieving performance improvements of up to 34.8%. These findings establish our method as an efficient new paradigm for the automated evolution of agent systems.

Abstract

Harness optimization provides a practical setting for recursive self-improvement (RSI), where agent-generated modifications inform subsequent changes through execution feedback. Recent work such as Meta-Harness implements this process through iterative code generation and evaluation, but retains a fixed development set and proposal policy. These constraints channel evolution along a single search trajectory, increasing the risk of converging to a local optimum. We make the improvement process itself adaptive by organizing search into branches with evolving development subsets and proposal policies. Each branch retains development cases solved by more of its leading harnesses than by those of other branches, drops cases solved by every leading harness across all branches, and revises its proposal policy using its own search history. To deploy the resulting complementary harnesses, we propose a router to select one development-selected branch head for each new input before execution. Across mathematical reasoning and agentic coding benchmarks, our system achieves relative improvements over Meta-Harness of 34.8% on Olympiad-level mathematical reasoning, 11.6% on Terminal-Bench 2.0, and 3.8% on SWE-bench Lite, with harness selection and router configuration based solely on development data. These results show that evolving branch objectives and proposal policies can yield complementary harnesses whose strengths a router combines without access to test outcomes.

Problem

• Recursive self-improvement
• Harness optimization
• Local optimum
• Search trajectory
• Agent

Innovation

• Recursive Self-Improvement
• Mixture of Branches
• Harness Optimization
• Adaptive Search
• Router Mechanism