All writing
RidingPlaybookOct 12, 2026 · 6 min

Your agent in a shared cabin: a rider's playbook

Grok Bot is now live in every Cybercab. Riding mode just gained an agent with your accounts, inside a car the next stranger will sit in ten minutes later.

On October 11, Tesla turned on Grok Bot in every Cybercab and the rest of its Model Y robotaxi fleet. A Tesla AI engineer announced it, and the company's Robotaxi account followed with a one-liner inviting riders to talk to their bot mid-ride. Early testers described the flow: sign in to Grok once, your bots are there on later rides, and the session logs out when you step out. The first demos are errands, like adding something to an Amazon cart or ordering a drink on the way to the coffee shop. The same weekend, Tesla said more than 300 Cybercabs are now serving Austin, roughly eight times the fleet it had a month ago.

That combination matters for work in motion. Riding mode used to mean a laptop and a connection. Now it also means an agent with your accounts, inside a car that isn't yours, that the next stranger will sit in ten minutes later. The work gets easier. The trust questions are new.

What changes when the cabin isn't yours

  • Your agent's session lives in a shared vehicle. Automatic logout on exit is the right default, but treat it like a hotel TV: confirm you're signed out before you close the door, and check your account's active sessions if anything looks off.
  • The cabin is observed. Tesla's rider guide says the Cybercab has a cabin camera by the dome lights, used for occupancy detection and support, and that a support representative may access it during a support call. The screen shows camera and microphone indicators. Glance at them before you say anything you wouldn't say in a meeting room with the door open.
  • Your voice is the interface. Anything you ask the agent is said out loud in the cabin. With a friend or colleague in the other seat, that's a shared call.

The Lego lesson: confirmations are credentials

Also this weekend, Elon Musk posted a screenshot of a Lego order his Grok Bot placed, with the order number, addresses, and the card's last four digits visible. Another X user typed that order number into Lego's own support bot and got a reply that a cancellation request had been sent. Musk said the order wasn't canceled. Whatever happened next, the pattern is the useful part: when agents on both sides of a transaction accept an order number as proof of identity, the order number works like a password.

  • Don't read order numbers, confirmation codes, or booking references aloud in a shared cabin, and don't screenshot them to share.
  • Ask your agent to confirm with a short summary, like "ordered, arrives Thursday," instead of reading the details back.
  • Keep payment methods stored with the merchant or your agent's account settings, never photographed into a chat.

How to use a robotaxi ride

As the Austin fleet grows, riders are reporting shorter waits, so the ride itself is becoming the predictable part of the trip. Plan by length:

  • Under 15 minutes: dispatch and errands. Hand off tasks by voice, let the agent run, and save the review for your destination.
  • 15 to 40 minutes: dispatch at pickup, then review on your own laptop or phone. Your device, not the cabin screen, is where you read drafts, check sources, and approve anything that sends.
  • Longer rides: a real work block. Close the loop in the car, but keep sends, payments, and shares behind an explicit "show me first."

A rider's setup checklist

  • Send agent output to a drafts folder by default, and put a spending cap or approval step on agent purchases.
  • Scope connectors to what you actually use on rides: your calendar and one work drive, not every account you own.
  • Keep sensitive work (client names, deal numbers, health, HR) on your laptop with headphones, not on the cabin microphone.
  • On exit, glance at the screen, confirm you're logged out, and take your device.
A shared cabin is a fine place to give an agent work. It's a poor place to read your secrets back out loud.

This builds on Cybercab as workstation, which asked whether the cabin would ever become a real work surface, and The handoff loop, which separates dispatch from review. This week's answer is partly yes: the agent arrived before the desk did. That makes the rider's job less about finding a work surface and more about running a clean session in someone else's car. Do that well, and a short Austin ride becomes ten minutes of work you didn't have to sit down for.

Get the weekly digest

One email a week: a playbook, a tool verdict, and what changed in autonomy.