Rivian's founder and CEO RJ Scaringe said at TechCrunch Disrupt this week that, in future, Rivian's cars will be able to run AI agents made by other companies, naming Anthropic as an example. "That's where it becomes magical," he said on stage, according to TechCrunch's report. Scaringe did not say when this would happen, and Rivian declined to share more, pointing to its "AI & Autonomy Day" event in December.
So this is a stated direction, not a product launch. It matters because it describes one way in-car AI could develop: the carmaker keeps the vehicle layer and lets the frontier labs' assistants plug into it. Here is what is known, what is not, and the safety line Scaringe drew.
TL;DR
| Question | Answer |
|---|---|
| What was said? | Rivian cars will eventually run third-party AI agents, such as Anthropic's |
| When? | Not stated; more expected at AI & Autonomy Day in December |
| Which agents? | Anthropic was named as an example; TechCrunch speculated about others, but Rivian named no partners |
| Safety line | "Anything that's vehicle-level controls are not being handed to an agent" |
| Today | Rivian Assistant, an in-house AI assistant on R1 vehicles (needs Connect+) |
| Architecture | In-house agentic framework that routes requests to local or cloud models |
What Rivian Assistant does today
Rivian's own support page describes Rivian Assistant as "your AI-based digital co-pilot." You activate it by saying "OK Rivian" or "Hey Rivian," by tapping an icon on the centre display, or by holding the left thumb control wheel. Listed capabilities include:
- Hands-free utility: switching drive modes, adjusting ride height, opening the front trunk, changing the temperature, or checking range-on-arrival estimates.
- Context-aware requests: for example, "warm up everyone's seats except mine."
- Navigation and media: directions, points of interest, and questions such as when a song came out.
- Messaging: reading, summarising and drafting replies to texts.
- Vehicle expert: answers owner's-guide questions, grounded, Rivian says, in "deterministic data, real-time vehicle logs and a custom built system for your Owner's Guide."
- General information: weather and local news summaries.
It is available in English on R1 vehicles in the United States and Canada, and requires a Connect+ subscription or trial. Turning it on switches off Alexa and Alexa integrations in the vehicle. Users must agree to AI Use Terms, and Rivian's footnote warns the technology "may occasionally provide inaccurate, incomplete, or outdated information." Memory, which lets the assistant remember things like a partner's name or favourite artists, is optional and disabled by default. TechCrunch adds that as of May the agent could reach external apps such as Google Calendar, which the support page now documents.
How the routing works, in Scaringe's words
Rivian says it built its own agentic framework that understands what action a person wants and routes it to the right part of the car's software. Scaringe described it this way: "Think of it as like a traffic light. So if you ask the vehicle something, it decides what model it's going to use. It can use a local model that's running on the edge and the vehicle. If it's a more complex question, it can go up to the cloud. And then even for that, we have different models that we use."
That is a familiar pattern: a small on-device model for latency-sensitive and offline tasks, and larger cloud models for open-ended questions. It matters for third-party agents, because plugging in an outside assistant means deciding which requests leave the car and what context goes with them. Our guide to always-on persistent agent architecture covers the same split between local memory and remote reasoning.
The hard part: plumbing, not prompts
Scaringe was candid that this is not a simple integration. "There's a lot of plumbing behind the scenes we have to do to make sure that that works, and it's not just within Rivian. That's also a lot of coordination with the labs to make sure they enable that plumbing," he said. TechCrunch notes that Rivian has built its in-car software from the ground up rather than relying on Apple CarPlay or Android Auto, which makes such integration possible but also means Rivian must define the interfaces itself.
He described the likely shape as an extension of how people already use agentic apps. A driver could ask for ventilated seats, then continue a conversation begun the day before at home about a Friday dinner, with the car pulling context from the conversation on another device. For that, Rivian "would have to find a way to connect the vehicle to all of the context" collected by apps such as Meta's Muse, per TechCrunch. The article also mentions OpenAI's "Dots" and Anthropic's Claude as examples of assistants Rivian might integrate; those were TechCrunch's illustrations, not announced partners.
The safety line
The most important statement is a limit. "You wouldn't want an agent to determine, 'hey, it's time to apply the brakes, or time to steer the vehicle.' So anything that's vehicle-level controls are not being handed to an agent," Scaringe said, adding that placing this software near safety-critical systems means Rivian has to be extremely careful.
That boundary is the right design principle for anyone building agents with physical consequences: separate the conversational layer, which can be probabilistic and replaceable, from the control layer, which must be deterministic and certified. It also leaves several open questions about the middle ground. Climate, seat heaters and the front trunk are low-risk, but what about door locks, charging limits or driver-assistance settings? Rivian has not said where it will draw the line beyond brakes and steering.
How this compares with other carmakers
TechCrunch notes that Rivian would not be the first American car company to add AI agents. Earlier this year, Tesla added a feature letting users talk to both SpaceX's Grok chatbot and its Grok Bot assistant; see our coverage of Grok Bot getting its own email address and Grok Voice 3.0. The difference in Rivian's pitch is that the lab's agent would be a guest inside Rivian's framework instead of the car's single built-in model.
This also touches a live fight over who gets to run agents on whose platform. Retailers have already blocked outside shopping agents, as in Amazon blocking Meta's Muse, and Meta's assistant is the one TechCrunch cites as a possible in-car guest; see Muse Wonder's launch details. A carmaker that opens its software to several labs becomes a gatekeeper of its own, deciding which agents get access to location, calendar and cabin context.
Privacy and security questions to ask before December
- What leaves the car? Voice audio, location and cabin context could go to a lab's servers. Rivian's existing terms cover its own assistant; third-party agents would bring their own terms.
- Whose terms apply? Rivian requires users to agree to AI Use Terms today. Will each third-party agent need separate consent, and can it be turned off independently?
- How are prompt-injection risks handled? Agents that read texts and calendar entries can be steered by hostile content in them. Experts have argued that labs often skip the basics here; see AI agents are not "going rogue". For teams building agent guardrails, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
- What about subscriptions? Rivian Assistant already requires Connect+, and it is unknown whether third-party agents would sit behind the same paywall, a separate lab subscription or both.
- Update and liability model. Over-the-air updates to a third-party agent could change behaviour without a Rivian release.
Why a carmaker might want guests instead of one model
There are practical reasons to open the slot. A single built-in assistant has to be good at everything: navigation, cabin control, general knowledge, writing and planning. Frontier labs spend far more on the open-ended part than a carmaker can, and their assistants already hold the context users care about, such as prior chats, documents and preferences. Letting those assistants appear in the car means Rivian can concentrate on what only it can supply: reliable access to the vehicle's data and a safe set of actions.
There are costs too. Each guest agent arrives with its own latency, its own privacy posture and its own failure modes, and the driver will blame the car for all of them. Rivian's "traffic light" routing is meant to keep the vehicle-specific requests (climate, seats, drive modes, range) on its own deterministic path, while open-ended conversation can go to whichever agent the user prefers. That split is also how the assistant stays useful when the connection is poor, because the on-board model can still handle cabin commands.
For owners, the near-term effect is limited: nothing changes until Rivian announces details. For other manufacturers it sets a template. Expect questions about whether a car should expose a standard agent interface, how apps get certified for the cabin, and whether a driver can bring their own assistant across vehicles in the way they bring a phone.
What to watch on December's AI & Autonomy Day
- A named list of lab partners and an interface description (is it a standard such as MCP or a bespoke SDK?).
- A permissions model: which car functions an agent can call, and which are blocked.
- Which processing happens on-device and which in the cloud.
- Pricing and whether owners can choose their preferred agent.
- Availability by region and vehicle generation.
What this means for builders
For developers, the signal is that automakers are treating the assistant as a swappable slot. If you build agents, expect more "host platform" environments with strict capability boundaries, where your agent receives a narrow tool set rather than full device control. Designing agents around small, well-described tools and explicit human confirmation is the portable skill.
Related reading
- How to build an always-on persistent AI agent
- Grok Bot gets its own email address
- Grok Voice 3.0 and Starlink
- Amazon blocks Meta's Muse shopping agent
- Meta Muse Wonder launch
- AI agents and security basics
Details are accurate as of October 15, 2026, from TechCrunch's Disrupt report and Rivian's support page; no date or partner list had been announced.