WANDR: an environment for evaluating agentic search
On July 14, 2026, Perplexity said it was making available an evaluation and reinforcement-learning environment for agentic search, synthesized from production traces with weak human supervision.
On July 14, 2026, Perplexity said it was making WANDR available, an evaluation and reinforcement-learning environment for agentic search systems. The company says the environment was synthesized from production traces with weak human supervision and is used internally to train models. The account does not detail evaluation metrics, results, or access terms.
The announcement is relevant to teams studying how to evaluate or train search systems; it does not, by itself, show that the method improves performance. Consult the original publication to confirm its scope and any technical details, and separately verify any results or terms of use that are disclosed. If using AI to study or apply the material, do not submit internal traces or documents without authorization: Rota Nacional detects personal data before execution and applies the organization’s policy.