WANDR Benchmark: Evaluating Research Agents That Must Search Wide and Deep

This post introduces WANDR, a new benchmark and evaluation harness for research agents. WANDR addresses the challenge of 'wide-and-deep' research by evaluating an agent's ability to discover a large set of entities and gather specific evidence for each. It defines a flexible qualification key hierarchy and a reference-free grading system. The post details the task construction pipeline and presents findings from evaluating six production systems, highlighting Perplexity's Search as Code as the leading performer.