Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows
Harness-Bench is a diagnostic benchmark designed to evaluate how different system layers, or 'harnesses,' affect AI agent performance in real workflows. It tests 106 practical tasks across various models and harness configurations. The findings show that agent success depends on the model-harness pairing, not just the model alone, revealing where reasoning breaks down when tools and workspace states are misaligned.