The Starvard sealStarvard.
English
EnglishFrançaisEspañolहिन्दीPortuguêsBahasa IndonesiaAll languages →
Starvard University

The Starvard Blog

Notes from the free university

AI Agents in VR: When Your Assistant Has a Body

October 18, 2026 · By the Starvard professors
A small holographic robot assistant on a kitchen counter pointing at a jar of chili crisp while a person in smart glasses cracks eggs
Illustration by Starvard (AI)

Your AI assistant has learned to talk, then to see, then to act on your behalf. The next step is the strangest one: giving it a body. This post separates the three kinds of body an AI can have, tours the glasses and headsets actually shipping in 2026, scores the hype claim by claim, and faces the hard problems, including what happens to the footage from a camera that shares your eyes. The short version: the agent part is real, the borrowed body is selling in the millions, and the visible companion on your kitchen counter is still a destination, not a product.

The kitchen thought experiment

Picture yourself standing in your kitchen, staring into the fridge, trying to figure out dinner.

Version one, 2015: you pull out your phone, type "recipes with leftover rice, two eggs, half an onion," scroll past four ads and a 2,000-word blog post about someone's grandmother, and eventually make fried rice.

Version two, 2023: you ask a chatbot. It gives you a decent fried rice recipe in eight seconds. Better. But you still had to type out what is in your fridge, because the chatbot has no idea. It lives in a box. It has never seen your fridge. It has never seen anything.

Version three, 2026: you are wearing a pair of glasses that look mostly like normal glasses. You say, "What can I make with this?" The assistant looks through a small camera at the same shelf you are looking at, notices the rice, the eggs, the onion, and also the jar of chili crisp you forgot you owned, and talks you through dinner while your hands are busy cracking eggs. Then, because you asked it to earlier, it reorders eggs.

Version four, coming soon-ish: a small, friendly figure is sitting on your counter. It points at the chili crisp. It watches you chop and says "smaller pieces, they will cook faster." When you laugh, it looks at you. It is not real, in the sense that there is nothing on your counter. It is real, in the sense that you can see it, it can see you, and it is doing things for you.

Each step in that sequence gives the AI one more thing it did not have before. First it got language. Then it got eyes. Then it got hands, the ability to take actions on your behalf. And the last step gives it a body: a visible presence in the space you share. That last step is what this post is about. Call it embodied AI, spatial AI or agents in XR; underneath the jargon it is one question: what happens when your assistant stops being a voice in a box and starts being something that shares your space?

What an agent is, and why the guardrails matter

A chatbot is a thing you talk to. Ask it to book a flight, and the best it can do is tell you how to book a flight. An agent is a thing that does stuff. You give it a goal, and it goes off and takes steps in the world to reach that goal: it opens websites, fills out forms, sends emails, compares prices, comes back and asks you a question when it gets stuck, and then keeps going. A chatbot answers. An agent acts.

This distinction stopped being theoretical in 2026. The clearest example is Muse, which Meta launched in the US on September 8, 2026. Meta describes it as a personal agent, and the feature list is very agent-shaped: it can send email, book travel, open a browser, fill out forms, negotiate on your behalf and complete purchases, and it keeps working in the background after you close the app. It runs on Meta's Muse Spark model, is available through its own app, the web and WhatsApp, and comes in a free tier with paid tiers at $20 and $100 a month.

What is interesting about Muse for our purposes is not the feature list. It is the safety architecture, because it shows what everyone building agents has had to confront. Muse runs inside what Meta calls a Muse Secure VM, a private cloud computer that holds the agent, your connected data and your credentials. A separate Sentinel agent watches what Muse sends out to the internet, an AI babysitting another AI. Muse can use your saved passwords and payment details without being able to see them. Sensitive actions, like sending an email or completing a purchase, require your approval first. And at checkout it uses Stripe's Link, including one-time card numbers; Stripe's own team called Muse the first AI agent covered by Link's buyer protections.

Why the guardrails? Because once software can act, mistakes get expensive. A chatbot that hallucinates gives you a wrong answer. An agent that hallucinates books you a nonrefundable flight to Portland, Maine instead of Portland, Oregon. Hold onto that thought. It gets much more important once the agent can see through your eyes.

What "VR" even means in 2026

Here is something that confuses almost everyone: "VR" in 2026 is not one thing. It is a spectrum of devices that differ mainly in how much of your vision they take over. At one end sit full headsets like Samsung's Galaxy XR, Apple's Vision Pro and Meta's Quest 3, which replace your view with screens while cameras pass the real world through. The Galaxy XR weighs 545 grams by Samsung's spec and costs $1,799; the Vision Pro is heavier still and costs several thousand dollars. One step down are the announced Meta VR Glasses, the same idea squeezed into a glasses shape with a pocket compute puck: about 100 grams on your face, per Meta, at $1,299.99, due spring 2027. Then come AR display glasses, see-through lenses with digital images layered on top: Snap's Specs at 132 grams and $2,195, and the Meta Ray-Ban Display at 69 grams and $799. Below those are AI camera glasses with no display at all, just a camera, microphones, speakers and an AI you talk to: the Ray-Ban Meta Gen 3 at $449 is the famous one. And at the far end, audio glasses like the Ray-Ban Meta Audio at $349 and 43 grams, with no display and no camera, only microphones, speakers and AI.

People argue about what counts as VR versus AR versus MR versus XR. For this post, XR (extended reality) is the umbrella for all of it, because the interesting question is not the label. It is this: the less a device covers your eyes, the more people actually wear it. That sounds obvious, but the numbers tell the whole story of where AI agents are going to live first.

EssilorLuxottica, Meta's eyewear partner, says it sold more than 7 million Ray-Ban and Oakley Meta glasses in 2025, more than three times the roughly 2 million sold across 2023 and 2024 combined. Meanwhile, analytics firm Treeview estimates that Samsung's Galaxy XR headset has sold around 70,000 units in its first year; that is an outside estimate, not a Samsung figure, but nobody credible is suggesting it is in the millions. And Counterpoint Research's data for the first half of 2026, as reported by MuyComputer, put 96 percent of AI-glasses shipments in the no-display category.

So the most successful "XR" device on Earth right now does not have a screen. It is a pair of sunglasses with a camera and a voice assistant. Notice the irony: the industry spent a decade trying to put a computer in front of your eyes, and what people actually bought was a computer that sits next to your eyes and looks where you look.

Three kinds of body

"Body" can mean three very different things, and most confusion about this topic comes from mixing them up. Think about what a body does for you. It gives you senses, a point of view: you see from somewhere, and you hear what is near you. It gives you presence: other people can see you, read your face, notice where you are looking. And it gives you hands: the ability to change the world, not just observe it. An AI can get each of these separately, so let's split embodiment into three kinds.

The borrowed body. This is the AI glasses model. The assistant does not have a body of its own; it rides along in yours. Its camera sits next to your eyes, its microphone next to your mouth, its speaker next to your ear. It sees what you see, from where you stand. This is what Google demoed live on the TED stage in April 2025, when Shahram Izadi, who leads Android XR, brought out prototype glasses. In the demo, a colleague turned away from a bookshelf, then a few seconds later asked the assistant for the title of a white book she had glanced at. Gemini told her. It had been watching along with her, and remembered. That moment is the borrowed body in a nutshell: the AI has your point of view, plus a better memory than you.

The visible body. This is the AI as a character: a face, a figure, something you can look at and that looks back. Game NPCs, virtual tutors, training simulations, a little hologram on your desk. The AI does not necessarily see through your eyes; it occupies a spot in a shared space, real or virtual.

The acting body. This is the agent part. The AI can change things: in a virtual world it can pick up objects and walk around; in your digital life it can send the email or buy the eggs.

Map today's products against those three and a pattern jumps out. The AI glasses (Ray-Ban Meta, Samsung's eyewear) have borrowed senses and growing hands, since Muse is coming to Meta's glasses and Google has demoed Gemini placing a DoorDash order in the background, but no visible body at all. The Galaxy XR with Gemini sees your passthrough view and acts across apps, still with no face. AI game characters built on NVIDIA ACE or Convai have visible bodies and hands inside their games, and perceive the game world rather than yours. VR training avatars have visible bodies but no hands of their own. And the kitchen companion from the intro, with all three at once, does not exist yet as a consumer product. Nobody has all three. The fully embodied assistant is the destination; today we have pieces. And here is the twist: the piece that is selling, by a huge margin, is the borrowed body, the one where the AI does not get a face at all. That might be telling us something about what people actually want.

Why would a body help, and why might it not?

Start with a question that sounds dumb but is not: why do you have a face? Not biologically. Socially. Humans read trust, attention and intent off tiny cues: where someone's eyes go, whether they nod, how close they stand. A huge share of human communication travels through channels a text box cannot carry. So the first-principles case for giving an AI a body goes like this. Bodies carry information: a pointing hand is faster than "the jar on the second shelf, to the left of the mustard," and a nod is faster than "I understand." Bodies create presence: being in the room with someone feels different from reading their messages, even if the words are identical. And bodies set expectations: if something looks like a tool, you use it like a tool; if it looks like a teammate, you treat it like one.

The case against is just as structural. Bodies are expensive: rendering, animating and lip-syncing a convincing character costs compute, battery and money. Bodies can be creepy: the closer something gets to human without quite making it, the more unsettling it becomes, the effect roboticist Masahiro Mori named the uncanny valley back in 1970. And bodies get in the way: you do not always want someone standing in your kitchen; sometimes you just want the answer.

So what does the research say? A sampling, with an upfront warning: most of these are small lab studies, often a dozen to a few dozen people. Treat them as interesting signals, not settled science.

A body can make you feel less alone. A study from Italy's National Research Council (CNR) gave people an AI tour guide in single-user VR in three forms: voice only, a human-sized 3D avatar, and a miniature hologram. With 12 participants, the visible versions significantly reduced feelings of being alone, and people split on whether they preferred the full-sized or mini version. Smaller might be better: a University of Florida study by Isaac Wang, Jesse Smith and Jaime Ruiz compared four AR assistants for real-world tasks (voice-only, non-human, full-size human and miniature human), and participants preferred the miniature embodied agent. A tiny helper on your desk, apparently, beats a full-sized stranger looming over you. Gestures build trust: work by Kim and colleagues, summarized in a 2023 paper on AI anatomy tutors in VR, found that an assistant that gestured and moved around was trusted more than one that did not. The kind of body changes how you judge it: a September 2026 study from KU Leuven built a VR puzzle game for Apple Vision Pro with an AI companion that could help in four ways (no help, following commands, discussing strategy, or playing autonomously); with 24 participants, AI help significantly reduced players' workload, but the revealing finding was qualitative: players judged the companion not just on usefulness, but on whether it felt like a tool, a teammate or a character in the world, and different players wanted different things. And sometimes people open up more to a virtual human: one of the best-known findings in this field comes from USC's Institute for Creative Technologies, where researchers led by Gale Lucas and Jonathan Gratch found in 2014 that people disclosed more sensitive information to a virtual interviewer when they believed it was fully computer-controlled rather than operated by a person. No human judging you, less holding back.

Put that together and a picture forms. Bodies help most when the task is social (practice a hard conversation, feel less alone, learn from a tutor), when the body is appropriately sized and styled (a mini helper, not a looming stranger), and when the body matches the role the person wants (tool, teammate or character). Bodies help least when you just need a quick answer while doing something else. Which is, not coincidentally, most of what people do with AI glasses.

How it actually works: the stack behind the magic

When you talk to an embodied AI, whether it is a character in a headset or a voice in your glasses, roughly six things have to happen, in order, fast. First, perceive: cameras and microphones capture the world. On a Quest 3, for example, Meta opened the headset's front color cameras to developers in March 2025 through the Passthrough Camera API, and the numbers give you a feel for the constraints: apps get 30 frames per second per camera at up to 1280 by 960 resolution. That is not a lot of pixels for understanding a whole room, and Meta's own docs warn that reading camera frames costs performance and battery. Second, hear: speech recognition turns your voice into text. Third, think: a language model reads what you said, what the camera sees, what it remembers about you and any instructions about who it is supposed to be, and decides what to say or do. Meta's open-source Spatial Scanner sample app shows the typical pattern: a small model on the headset spots objects, then sends a picture to a big vision model in the cloud for a detailed answer. Fourth, act: the agent calls tools, whether that is search, calendar and checkout, or, in a game, "walk to the door." Fifth, speak: text-to-speech gives the response a voice. Sixth, embody: if the AI has a visible body, its face and body need to move in sync with that voice, which is what NVIDIA's Audio2Face, part of its ACE toolkit, generates from audio.

Now here is the catch that makes this whole field hard: all six steps have to happen in less time than a human pause. A July 2026 developer guide from Cinevva summarizes the research this way: responses under roughly 800 milliseconds feel conversational, while a typical round trip to a cloud AI adds one to two seconds of lag, which reads as the character stalling. A one-second pause from a face that is looking right at you feels like talking to someone who is not quite there. That latency problem pushes you toward running the AI on the device. But running big models on the device runs straight into the next problem: power and heat on your head.

Look at how every company is solving it. Samsung's Galaxy XR uses a tethered battery pack and gets about 2 hours of general use, per Samsung. Meta's VR Glasses move the chip, battery and storage into a pocket puck connected by a cable, which is how Meta gets the part on your face down to about 100 grams; early reports say up to about 3 hours of video. Samsung's Intelligent Eyewear offloads the demanding AI work to your phone, and reported battery life is around 9 hours, because the glasses themselves are not doing the heavy thinking. Snap's Specs are fully standalone with two Snapdragon chips in the frames, up to about 4 hours of mixed use by Snap's claim, and early hands-on reports noted the demo units were heavy and could run warm. Physics wins every argument here. The more an AI wants to see and render on your face, the more battery it eats, the hotter it runs, and the heavier it gets. That is a big part of why camera-and-voice glasses, with no display, no avatar and the AI in the cloud or on your phone, are the ones people wear all day.

The 2026 landscape: who is actually doing this

Meta, all-in and expensively. Start with the money. Meta's Reality Labs division lost $4.62 billion in Q2 2026 on revenue of $431 million, per Meta's earnings report. It lost about $19.2 billion in 2025, and cumulative operating losses since late 2020 are over $80 billion. For a sense of scale: in a single quarter, Reality Labs spent roughly eleven dollars for every dollar it brought in. The pivot matters more than the losses. In January 2026, Meta cut hundreds of Reality Labs jobs and moved Horizon Worlds, its flagship social VR app, into maintenance mode. CFO Susan Li has signaled that VR spending is coming down while spending on glasses and wearables goes up, and one analyst firm reports that around 70 percent of Reality Labs operating expenses now target AI glasses and wearables. The metaverse has been quietly retired in favor of AI on your face.

At Connect 2026, on September 23 and 24, Meta announced the Ray-Ban Meta Gen 3 at $449 with a new action button for the AI; the Ray-Ban Meta Audio at $349, its first glasses with no camera at all, at 43 grams, a direct answer to privacy concerns; a plan for more than 100 glasses styles by the end of 2026; and the Meta Ray-Ban Display, $799 and 69 grams, with a small 600 by 600 display in the right lens controlled by a wrist-worn Neural Band that reads muscle signals, expanding to Canada, the UK, France, Italy and Germany after months of being sold out in the US. Muse is coming to the glasses in the coming months, activated by saying your agent's name, and Meta also previewed the Muse Charm, a small Tamagotchi-like pocket device that gives you constant access to your agent without glasses at all. And then the most body-relevant announcement: the Meta VR Glasses, previously codenamed Project Phoenix. $1,299.99, about 100 grams on your face, a pocket puck with a Snapdragon chip, a 5K micro-OLED display, controlled with eyes, hands and voice, with Meta AI built in, on sale spring 2027 with no preorders yet. Early hands-on reports put the field of view around 70 degrees, noticeably narrower than a Quest 3. The glasses also ship with photorealistic hologram calling built on Meta's Codec Avatars research, which UploadVR reports starts as a 2.5D stereoscopic video stream with full 3D planned later. That matters for our story: Codec Avatars are exactly the technology you would need to give an AI a lifelike face. Today they represent people. It is not hard to guess where else they could go.

Google and Samsung: the Gemini bet. If Meta's pitch is "AI on your face," Google's is "Gemini everywhere, including your face." The Galaxy XR launched in October 2025 at $1,799 as the first device running Android XR, the platform Google built with Samsung and Qualcomm; its headline AI feature is Gemini Live, which can see what you are seeing in the headset and act across your apps. On the glasses side, at Google I/O in May 2026 Google said there would be two kinds of Android XR glasses: audio glasses that talk in your ear and display glasses that show you information. Samsung's first model, built with Gentle Monster and Warby Parker, has no display and is launching in November 2026, a Samsung official told The Korea Herald. It uses a camera, microphones and speakers, and hands the heavy AI work to your phone. Korean media report it weighs about 46 grams, five less than a base Ray-Ban Meta, with LED indicators inside and outside the frame that light up when it is recording; the price has not been announced, and numbers floating around online are guesses. Google also previewed Android Halo, a little strip at the top of your phone screen that shows what your agent is working on at any moment. Small detail, smart idea: when an agent acts in the background, you need a way to glance at what it is doing.

Apple: polished, pricey and patient. The Vision Pro got its first Apple Intelligence features with visionOS 2.4 in March 2025, and at WWDC in June 2026 Apple announced visionOS 27 with a more capable new Siri for the headset, arriving as a free update this fall. Apple has not shown anything like a visible AI character, and at its price the headset remains a device very few people own. Bloomberg's Mark Gurman has reported that Apple's own AR glasses are not expected before 2028 at the earliest.

Snap: the long-shot AR bet. Snap unveiled Specs, its first consumer AR glasses, at the Augmented World Expo on June 16, 2026: $2,195, shipping this fall in the US, UK and France. They are fully standalone, with no puck, weigh 132 or 136 grams depending on size, and have a 51-degree field of view. Developers can build with multimodal models from OpenAI and Google, and Snap's pitch for its SPECS Intelligence is anticipatory help: assistance before you ask, but no actions without your yes.

The character builders. The visible-body AIs come from a different set of companies: NVIDIA's ACE toolkit, Convai and Inworld, whose characters perceive, converse and act inside games and virtual worlds. We toured that workforce in depth in AI Agents in VR: Virtual Worlds With a Real Workforce, so here it is enough to note the honest state of play, per that same July 2026 Cinevva guide: ACE is best thought of as voice-and-animation middleware rather than a complete character brain, Inworld has largely pivoted from a plug-and-play character studio toward B2B voice and AI infrastructure, and Convai, with a free tier and a $29 a month indie plan, has become a common default for developers.

The trainers. The most mature use of AI with a visible body might be the least glamorous: workplace training. Mursion, founded in 2015, builds VR simulations where you practice hard conversations (giving feedback, handling an angry customer, delivering bad news) with avatars. Here is the honest part that is easy to miss: Mursion's model has long blended AI with live human simulation specialists who orchestrate the avatars behind the scenes. In other words, some of the most convincing "AI characters" in production have had a human in the loop. That is not a scandal; it is a clue about how hard this is. Bodyswaps takes a more automated approach, with AI avatars and AI-driven feedback for soft skills in VR or on a screen, including its project with the Royal Society of Medicine aimed at clinicians.

The companions. At CES 2026, Razer showed Project AVA, a prototype 3D avatar AI assistant that appears inside a desktop device, expected in the second half of 2026, and Lepro showed Ami, a desk companion with a curved screen and eye-tracking cameras, which its maker explicitly markets as an "AI soulmate." Researchers are exploring stranger territory still: in a 2026 paper called CyberSelf, people talked in VR with an AI version of themselves, with their own face and cloned voice.

Hype versus reality, scored honestly

"Smart glasses are going mainstream." Verdict: partly true, and more than skeptics expected. Seven million-plus Meta glasses in a year is real. EssilorLuxottica originally targeted 10 million units of annual production capacity by the end of 2026, and Bloomberg has reported discussions of going to 20 million or more. But keep it in proportion: Samsung and Apple each sell over 200 million phones a year. Glasses are a real product category now. They are not yet a phone replacement.

"AI glasses give you an assistant that sees the world." Verdict: true, with an asterisk. This works today, on millions of faces. The asterisk is reliability. The most famous moment of Connect 2025 was Mark Zuckerberg failing to answer a WhatsApp call on stage with the brand-new Ray-Ban Display. Google's TED demo had a small connection delay. The real world is messy, and these systems are still learning it.

"Your assistant will have a face and live in your room." Verdict: mostly not yet, for consumers. Visible AI characters exist in games, training sims, research labs and early companion gadgets. But the market is voting for faceless: 96 percent of AI-glasses shipments in the first half of 2026 had no display at all, per Counterpoint as reported by MuyComputer. You cannot show an avatar on glasses with no screen.

"VR headsets are the future of AI." Verdict: the industry itself has backed away from this. Meta, the biggest VR company, put its flagship social VR app in maintenance mode, is cutting VR investment and is pouring money into glasses. Its new "VR" device is literally called VR Glasses. The headset is not dead (it is great for games, movies and focused work), but it is not where most people will meet their AI.

"The XR market is worth hundreds of billions." Verdict: be very skeptical of numbers like this. In early 2026, one widely circulated article cited a Research Nester estimate putting the XR market at $336.5 billion in 2026. Compare that to the actual reported revenue of the biggest XR company on Earth: Reality Labs brought in $431 million in its most recent quarter. Figures like this usually fold in huge adjacent categories and vary wildly between firms. Even honest market-share numbers depend on who is counting what: an IDC figure reported secondhand gives Meta about 69 percent of the overall smart-glasses market in Q1 2026, while Counterpoint data has Meta at about 94 percent of the no-display AI glasses segment. Both can be true; they are measuring different things. When you see a giant number, ask: what exactly is being counted, and who is selling it?

"AI characters will replace scripted game characters." Verdict: experiments, not yet the norm. The tech works in demos and a handful of shipping titles. But latency, cost per conversation, guardrails (nobody wants the medieval innkeeper explaining cryptocurrency) and plain quality are unsolved at scale. As one tester put it after talking to NVIDIA's Jin, the characters were impressive but stilted, and good writing still matters.

"Agents will handle your life for you." Verdict: the software is arriving faster than the trust. Muse, Gemini's background tasks and Snap's anticipatory AI are all real and launching. Every one of them is wrapped in approval gates, because companies know people are not ready to hand over the keys unattended. Whether people will pay $100 a month for an agent is a genuinely open question.

The hard problems nobody has solved

Problem 1: an assistant that sees everything you see shows it to whom, exactly? Remember the borrowed body: the AI sees what you see. That is the feature. It is also the problem. In February 2026, the Swedish newspapers Svenska Dagbladet and Göteborgs-Posten published an investigation reporting that footage from Meta's AI glasses could be routed to human data-labeling contractors at a firm called Sama in Nairobi, Kenya. Workers described seeing deeply private moments captured inside people's homes, including people undressing and using the bathroom, as well as bank card details. Meta's response was that it sometimes uses contractors to review data shared with Meta AI in order to improve the product, as many companies do, and that it filters the data to protect privacy. A US class-action lawsuit followed, focused on Meta's privacy marketing, and the UK's Information Commissioner's Office said it was contacting Meta. Why does this matter for embodied AI specifically? Because an AI that shares your body shares your life: your bathroom, your bank card, your kids. "It is in the terms of service" will not be a good enough answer. You can see the industry reacting: Meta launched camera-free audio glasses and is working on private processing for glasses; Muse runs in an isolated VM, and Meta has said a Confidential VM encrypted with a key only you control is coming; Samsung is adding recording lights on both sides of the frame. These are reasonable steps. None of them settles the question.

Problem 2: the people who did not agree to any of this. You chose to wear the glasses. The barista did not. An AI that sees through your eyes is also seeing them. Recording LEDs and glasses-detector apps help a little, but the social norms are still forming, much as they did when phones first got cameras.

Problem 3: when the agent acts, who is responsible? A chatbot's mistakes are words. An agent's mistakes are purchases, emails and appointments. An embodied agent's mistakes could be worse: imagine it misreading a label, a medication bottle, a street sign. The current answer is "ask the human first," which is sensible but caps usefulness. An assistant that asks permission for everything is just a slower you.

Problem 4: bodies make us bond. This is the subtlest one. Research suggests visible bodies reduce loneliness and increase trust. Those are features in a tutor. They are more complicated in a product designed to keep you engaged, sold by a company with an advertising business. (To its credit, Meta says Muse conversations are not shared with its ad systems.) An "AI soulmate" is a very different thing from a hint button in a puzzle game, and we do not yet have norms for how persuasive a friendly face should be allowed to be.

Problem 5: physics. The dream device (light as glasses, sees everything, renders a lifelike companion, lasts all day, no lag) does not exist. Every company is choosing which compromise to make first.

So where is this actually going?

Near term, now to about 2028: the body-borrower era. The dominant form of "AI with a body" will be the one already winning: an agent that borrows your senses and acts on your behalf, with no face at all. Voice in your ear, camera by your eye, maybe a small display. Muse coming to Meta's glasses, Gemini on Samsung's eyewear, Snap's anticipatory assistant. The battleground: does it get things done, does it handle your data well, and can you wear it all day?

Medium term: bodies where bodies pay off. Visible AI bodies will keep growing in places where social presence is the whole point. Training and practice: rehearsing hard conversations, medical communication, interviews. Games: companions and characters that respond to you, as tools like ACE and Convai mature and costs drop. Learning: tutors that can point at the anatomy model instead of describing it. And calls: Codec Avatar-style photoreal presence for people first, with AI characters borrowing the same technology later.

Longer term: the kitchen companion. The full vision from the start of this post, an assistant that sees what you see, has a visible presence beside you and acts on your behalf, needs several things to all happen at once: display glasses light enough to wear all day, AI fast enough to respond within a human pause, privacy architecture people actually trust, and social norms that make it okay. Each of those is moving. None of them is done. My honest guess: when that companion arrives, it will be smaller than you expect (remember the research preferring the mini helper), it will stay out of the way until you need it, and it will feel less like a person and more like a really good tool with a face. The lesson of the last three years is that people want their AI present, but not in the way.

The big picture

Let's go back to the kitchen one last time. AI has been getting closer to us: first it learned to talk, then to see, then to act. Now the question is whether it should also learn to be there, visible and sharing your space. The market's first answer has been "be there, but don't make me look at you." The best-selling embodied AI on the planet is a voice that looks where you look.

That could change; displays get lighter and humans love faces. But for now, the most honest summary of AI agents in VR in late 2026 is this. The agent part is real and arriving fast. The borrowed-body part is real and selling in the millions. The visible-body part is real in games, training and labs, and still early everywhere else. The fully embodied companion is a destination, not a product. And the hardest problems are not technical. They are about trust. When your assistant has a body, it isn't just a better assistant. It is a new kind of relationship with a machine, and a new kind of relationship with the company that made it. Worth thinking about before you put it on your face.

Why this belongs at Starvard

An assistant with a body sits at the crossing of three Starvard subjects: how AI actually reasons, how machines perceive the world, and the code that ties the two together. If this post left you curious about the machinery, start with How AI Works, our plain-English tour of what a model is and does. Go deeper with Machine Learning 101, which covers how systems learn from data, the same perception and memory loop the glasses in this post run all day. And if you would rather build a small character of your own, Python 101 is the friendliest first step toward the APIs every company in this post exposes. All three are free, like everything at Starvard.

Watch

Sources

Tagged: Technology · AI Agents · Virtual Reality · AI Agents in VR · Virtual Worlds

Back to the blog

A note on accuracy: Much of what you read at Starvard was created with help from AI agents, working alongside people. We check what we can, but we cannot guarantee that everything here is true, current, or complete. If something really matters, your health, your money, a grade, a job, please verify it with a trusted source before you rely on it.