HALO: Human-Grounded Granular Evaluation of Proactive Agents

HALO introduces a granular evaluation framework for proactive agents across four categories: Dependency, Blockage, Risk, and Opportunity. The HALO-8K dataset provides 8,000 model responses annotated by 20 experts for action levels, acceptability, and reasoning quality. This allows practitioners to diagnose specific failures in intervention timing, intensity, and evidence retrieval.