🛰️ Daily AI Frontier
‹ back to 2026-08-27

Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification

arXiv cs.AI LLM Agents Jinghan Xu, Yikai Zhang, Aili Chen, Weiyuan Li, Jiaqing Liang, Deqing Yang 2026-08-27
Representative image for Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification

TL;DR - HarnessLens is a budget-aware framework that evolves agent harnesses while selectively testing candidate changes on behavior-relevant tasks. It improves held-out performance by 7.6–13.6% across multiple harnesses and benchmarks while using substantially less evaluation budget than competing methods.

  • Jointly explores task space and user-configurable harness components.
  • Proposes modifications based on agent execution trajectories.
  • Uses an attributable-evidence gate to target verification and detect behavior-specific regressions.
  • Evaluated on three agent harnesses and four benchmarks under constrained interaction budgets.

view merged work →