WANDR Benchmark Tests AI Agents on Wide-and-Deep Research Tasks

WANDR is an open benchmark and evaluation harness built around 500 realistic data-collection tasks for knowledge work. Unlike its sibling DRACO, which evaluates long-form report writing, WANDR measures whether an agent can build a large, open-ended collection of entities and back every member with verifiable evidence. Tasks mirror real workflows like competitive mapping, due diligence, and literature review, using composable hierarchies such as company-employee-URL chains. Even the strongest system tested reached only 0.363 soft F1 and 0.133 hard F1, showing that combined wide-and-deep research remains far from solved. Grading is reference-free: instead of a fixed answer key, the system re-fetches each cited URL and verifies that excerpts and pages actually support every claim. This lets engineers localize failures to specific stages, whether discovery, enrichment, identity resolution, or evidence extraction, rather than treating research quality as a single opaque score.