MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution
Zhiwei Shang, Bryan Kian Hsiang Low (NUS), Chenglin Wu (DeepWisdom), Zhongxiang Dai (CUHK-Shenzhen) and colleagues propose MESH-Harness, which splits an agent harness into modules with fixed interfaces and searches over module combinations with a contextual bandit under a limited evaluation budget.
Ask this paper
Modules. Each harness is split into functional modules with role-specific interfaces, so alternative implementations of a module can be swapped and recombined. Validation traces guide local code edits that add new implementations to fixed-size pools.
Search. Full-covariance LinUCB over shared module representations scores combinations by predicted performance and exploration value, and mixed-start coordinate ascent picks configurations without enumerating the combinatorial space.
Results. Under matched evaluation budgets MESH-Harness beats Meta-Harness by 5.70 points on text tasks, 7.01 on retrieval-augmented math, 2.00 on LiveCodeBench and 5.00 on ScienceWorld, and improves 5.63 to 7.79 points over its own first round.
Cost. Test-time cost of the reported configurations is 44.2% lower than Meta-Harness and total cost including search is 14.6% lower. Random module selection scores 7.53 points below the full method.
Abstract
An agent harness is the code that organizes context, maintains state, and coordinates tool calls for a language model. We study how to improve the harness under a limited evaluation budget while keeping model weights fixed. Our method, MESH-Harness, organizes each harness into functional modules with explicit role-specific interfaces, allowing alternative implementations of each module to be substituted and recombined. It uses shared module representations and full-covariance LinUCB to score candidate combinations based on predicted performance and exploration value. Mixed-start coordinate ascent selects complete configurations for evaluation without enumerating the combinatorial space. Validation traces then guide local code edits, and the resulting candidates are incorporated into fixed-capacity role-specific pools for subsequent recombination. On text tasks, retrieval-augmented mathematical reasoning, code generation, and interactive scientific tasks, MESH-Harness outperforms Meta-Harness by 5.70, 7.01, 2.00, and 5.00 points, respectively, under matched candidate-evaluation budgets. Iterative harness optimization improves MESH-Harness by 5.63-7.79 points over its first-round configurations. For the reported configurations, aggregate test-time cost is 44.2% lower than that of Meta-Harness, while total cost including search is 14.6% lower. These results show that combining module-level design reuse with feedback-driven compositional search can systematically improve agent harnesses while keeping overall optimization cost under control.