Perplexity AI Launches WANDR Benchmark for Research Agents

Perplexity AI released WANDR (Wide ANd Deep Research) this week, an open benchmark and evaluation harness designed to assess the capabilities of AI research agents. This new benchmark addresses a gap in existing evaluations by focusing on the creation of large, evidence-backed collections of data rather than single answers. WANDR is positioned as the "wide" counterpart to Perplexity's previous DRACO benchmark, which focused on deep, long-form report generation.

WANDR evaluates agents on their ability to meet two key demands simultaneously: "wide" discovery of a broad set of qualifying entities, and "deep" investigation of each entity to support claims with evidence. This dual requirement means that agents must go beyond a few compelling examples or polished narratives based on incomplete research. The benchmark employs a composable qualification key hierarchy to structure these demands, allowing for tasks that can represent flat lists, nested searches, or matrices.

For instance, a task within WANDR might require an agent to identify a specified number of qualifying companies, then a certain number of employees for each company, and finally a set of supporting web pages for each employee. Each complete path through this hierarchy is validated independently. This structure ensures that agents are tested on their thoroughness and ability to gather comprehensive, verifiable information.

A concrete example provided by Perplexity is the `ceo_cfo_appointments` task. This task requires agents to identify at least 70 US-based companies where a CEO or CFO appointment was first announced between March 1 and April 30, 2026. For each company, the agent must supply one authoritative announcement page and a separate listing-authority page. This specific task necessitates the collection of 140 source-backed records, demonstrating the benchmark's focus on detailed, evidence-driven data acquisition.