Perplexity AI Releases WANDR: Open-Source Benchmark for Evaluating Agent Breadth and Depth

Perplexity AI has launched WANDR (Wide ANd Deep Research), an open-source benchmark with 500 realistic data-collection tasks for knowledge work. It extends their earlier depth benchmark DRACO to test whether agents can find many qualifying entities (breadth) and provide verifiable citations for each (depth). Tasks use a composable hierarchy of eligibility keys—e.g., require at least 70 US companies, each with a CEO/CFO appointment announcement and an authoritative source page, totaling 140 records with sources. The evaluator re-fetches submitted URLs to verify that excerpts are present and support all requirements. Perplexity’s Search as Code system achieves a soft F1 of 0.363 and hard F1 of 0.133 on all 500 tasks, leading Anthropic (0.249 soft F1) and others. Costs vary widely from $0.03 to $324.83 per task. Key findings: all systems have lower recall than precision; deeper hierarchies cause more failures; top-level entity discovery is the primary bottleneck; finding accessible pages is often easy, but extracting full evidence is hardest—Perplexity misses critical info in 41.4% of pages. For teams building market analysis, due diligence, or talent search agents, WANDR offers a realistic evaluation framework. It not only measures overall performance but also pinpoints failure stages (discovery, completion, or evidence extraction) via a scoring tree, guiding targeted improvements. The benchmark code is open-source for developers to run their own evaluations.