Mistral launches Agentic Search for complex documents
Two moves today show AI shifting from flashy demos to the unglamorous plumbing that actually makes models useful: a retrieval layer that reads inside dense filings, and Tesla quietly pulling humans out of its robotaxis.
Mistral's new Agentic Search turns RAG from a single guess into a multi-step investigation. The company launched Agentic Search on August 20, a retrieval layer that lets an AI system navigate, open, read, and verify information inside long, table-heavy documents instead of answering from a fixed set of retrieved text chunks. Where classic retrieval grabs the top-k passages and hopes the answer is in them, Mistral's loop gives the model five file-system-style tools — search, open, navigate, read, and grep — so it can drill into a specific page or table, follow a reference, and confirm the evidence before responding.
The numbers Mistral cites are large enough to matter. On FinanceBench, a question-answering benchmark over 368 SEC filings averaging roughly 147 pages each, moving from one-shot retrieval to the agentic loop lifted accuracy about threefold (26.7% to 86% on its internal measure), and adding the navigation tools trimmed p90 latency by up to 39.6% while cutting token use by as much as a third. On OfficeQA Pro, a harder benchmark of scanned Treasury bulletins, the full loop reached 51.9% accuracy versus 6.3% for one-shot retrieval. Because the tools need no fine-tuning, Mistral argues search quality should keep improving as underlying models get better — retrieval stops being capped by your chunking strategy and starts scaling with reasoning ability. It's available through the Mistral Search Toolkit and built into Studio and Vibe.
Tesla's Austin robotaxis appear to have gone fully driverless. Over the past two weeks, every one of the 170 Tesla Robotaxi rides in Austin tracked by the crowdsourced Robotaxi Tracker ran with no human safety monitor onboard, the site's creator, Ethan McKanna, told The Verge — spanning 54 different cars. McKanna also reports roughly 30 unsupervised Teslas operating in Dallas and Houston over the past week, a sharp reversal from earlier this summer when his tracker showed the unsupervised fleet shrinking. Some of the jump is better data: McKanna, who interned with Tesla's Robotaxi team this summer, says lagging metrics on his site undercounted real activity, and he is revising the sources feeding the tracker. But the Austin shift itself looks real — plates previously marked supervised have been confirmed offering unsupervised rides, and rider reports on Reddit and X corroborate it.
The timing is deliberate. Tesla is preparing to launch the Cybercab, a purpose-built robotaxi with no steering wheel or pedals, in Austin as soon as this month — a vehicle that only makes sense if Tesla can run without onboard monitors. Even after the jump, Tesla trails its main rival badly: Waymo runs more than 300 driverless vehicles in Austin alone and around 4,000 across eleven cities. Tesla did not respond to requests for comment, so the unsupervised status rests on tracker data and rider anecdotes rather than a company confirmation.
What to watch: whether Tesla publicly confirms the unsupervised status before the Cybercab launch, or keeps letting crowdsourced trackers do the reporting for it.
Should robotaxi operators have to disclose their supervision status, or is silence just part of the race? Tell us in the comments.
