VeriHarness turns a large language model into an agentic verifier equipped with workspace tools and reusable verification skills for checking long-horizon agent outputs. The system combines a disagreement resolver, which checks competing claims against environmental evidence, with a consensus challenger that stress-tests claims multiple agents agree on. Across five benchmarks and two frontier models, evidence-backed revision with VeriHarness improved selection scores by 6.2 to 6.4 points over single rollouts. The authors also show its verification skills can self-improve from failure feedback, offering a way to scale verification for long-horizon agentic tasks.
