Rivian Plans Outside AI Agents in Its Cars, Workday Makes Sana Its Home Screen, and Gemini Tasks Keep Running After You Leave

The chat box is quietly becoming the front door to other people's software. This week brought three examples: a carmaker preparing to host assistants it did not build, an enterprise vendor replacing its home screen with a conversation, and Google's assistant working on after you close the window. The engineering question underneath all three is the same: what is the agent allowed to touch, and who finds out?

1. Rivian Wants to Host Agents It Did Not Build

At TechCrunch Disrupt on October 13, Rivian CEO RJ Scaringe said the company plans to let outside AI agents run in its vehicles. No launch date was given, and Rivian pointed to its December AI & Autonomy Day for more. Coverage mentions Anthropic as an example of a provider that could be involved, and some outlets speculate about others, but Rivian has not confirmed any partners.

The detail worth noting is the boundary. Per reports of the interview, Scaringe said controls such as braking and steering will never be handed to an agent. Rivian's current in-house assistant already routes requests between a small on-device model and cloud models depending on complexity, and has reached apps like Google Calendar since May.

That is a useful template for anyone exposing a product to third-party agents: decide up front which capabilities are structurally unreachable, rather than relying on a prompt to say no.

2. Workday Turns the Assistant Into the Home Screen

On October 13, at Workday Rising, Workday described Sana as the new Workday Home, an agentic experience that pulls tasks, decisions, meetings and agent activity into one place. Employees ask for an outcome in natural language and Sana orchestrates Workday agents to complete it, with a Concierge handling HR and IT requests. Earlier this year Workday said Sana connects to hundreds of enterprise applications. These are vendor claims, and we could not reach the press release directly, so treat the specifics as company positioning.

The design choice matters more than the feature list. When the conversation is the default landing page, every menu a user used to click becomes a tool the model must pick correctly. Tool descriptions, argument validation and clear failure messages stop being polish and start being your navigation.

3. Gemini Tasks Keep Working Without You

Google's support documentation says Gemini can work through multi-step tasks in select Android apps in the background, even after you close the chat, with notifications and the ability to step in or stop. Android Authority also reported in September that a Tasks mode was starting to appear in the desktop Gemini app. Treat anything beyond the documented Android behavior as early.

Background work changes the contract with the user. A reply is no longer the end of a turn, so you need status, interruption and a record of what was done while nobody watched. A minimal shape for that:

type TaskEvent = {
  taskId: string;
  tool: string;
  action: 'started' | 'needs_approval' | 'done' | 'stopped';
  at: string;
};

async function runStep(taskId: string, tool: string, risky: boolean) {
  await log({ taskId, tool, action: risky ? 'needs_approval' : 'started', at: new Date().toISOString() });
  if (risky) return waitForUser(taskId); // never run it silently
}

What Connects the Three

Each story moves the chatbot from answering to acting inside someone else's system: a car, an HR suite, a phone's apps. The shared lesson is that the permission model is the product. Suggested visuals: a diagram of a car with an "agent-reachable" layer versus a locked driving layer, and a timeline of a background task with approval points marked.

Conclusion

If your chatbot can act, write down what it can never do, log what it did while unattended, and make stopping it one tap. This week's announcements all assume users will ask for those.

— Maya