Mistral released Agentic Search on 20 August as a retrieval layer that lets a model navigate, read and verify information inside complex documents rather than retrieving a passage and hoping. It introduces a multi-step loop for finding, inspecting and verifying information across data sources, and Mistral reports more accurate results while reducing turns, token use and latency against the FinanceBench and OfficeQA Pro benchmarks. It ships through the Mistral Search Toolkit and is built into Libraries in both Studio and Vibe.
The counterintuitive part is the cost direction. A retrieval loop with more steps would normally spend more tokens; Mistral's claim is that verifying inside the document reduces total turns, because the model stops re-asking when the answer it got was thin. The other stated property is isolation: portable and open tooling so enterprises can use sensitive domain-specific data without crossing their isolation boundaries in cloud or on-premises deployments.
Separating the operations is the design this site argued for in agents searching, fetching and browsing as distinct operations, and the benchmark question it raises is the one from two ExtractBenches measuring different systems: a retrieval score depends on which part of the pipeline is being scored. The agents hub tracks that boundary.
The accuracy, turn and latency numbers are Mistral's own against two named benchmarks, without a published per-benchmark breakdown or baseline configuration. The condition to watch is whether the turn reduction holds on documents the vendor did not choose, because a verification loop that saves turns on FinanceBench can still thrash on messier corpora.