Skip to main content
🤖 Running agents for a team, not just yourself?Get an independent review of identity, secrets, failover, observability and governance.Assess your agent platform
AI

On-Device AI With Div Garg of AGI, Inc.

I sat down with Div Garg, founder and CEO of AGI, Inc., on why the future of AI is on-device: privacy, ownership, and agents that act for you.

Luca Berton

· The Production AI Expert

Published ·5 min read

I sat down with Div Garg, the founder and CEO of AGI, Inc., for a conversation about the future of intelligence — and why it is moving off the cloud and onto your devices. Div built one of the first agent systems in the world: he founded MultiOn, the first company to deploy computer-use agents in the browser, and is an adjunct CS professor at Stanford. AGI, Inc. (founded in San Francisco in 2025) is his next bet: on-device intelligence that can act for you, privately, on the hardware you already own.

“The biggest problem with today’s AI is that it’s costly. To run a large model you need a big NVIDIA GPU in the cloud, and then you pay a token cost for every call. That makes it hard to scale to everyone cheaply. On-device changes the math.”

The problem with renting intelligence in the cloud

Almost every personal AI experience today is a cloud service: your data goes out, a big model runs somewhere else, and you pay per token. Div’s argument is that this model caps how far AI can go. If intelligence has to live on rented infrastructure, it can never be truly cheap, truly private, or truly available to the next billion people.

AGI, Inc.’s pitch is the opposite. Run the model on your device — or on your own cloud — and the cost structure flips. No per-call fees, no round-trip to a data center, and the data never leaves. That is the through-line of everything Div is building.

From renting an agent to owning one

The part that stood out to me was the shift from renting to owning.

“Instead of renting an agent from someone else, we want companies and users to be able to own it over time — and run the models themselves, on their own devices or their own cloud.”

The current “new wave of personal agents,” as Div put it, quietly hands over control: you paste in your data, you stop owning it, and you depend on whoever hosts the model. AGI’s thesis is that the agent should be yours — privacy-first, with your data staying on your side. It is a meaningful reframe: not “your AI assistant lives in the cloud,” but “you own the intelligence and where it runs.”

Div Garg on why on-device AI means you own your agent, not rent it

What AGI, Inc. actually ships

For context on what this is today: AGI, Inc. does original research on models that act, and ships them as products — AGI-ACT (a small action model), AGI Agent (which works across any app), and AGI-RUN, an inference engine that runs the stack at the edge. The agent is described as “computer-use”: it sees the device screen, understands context, and takes actions across any application — booking a restaurant, ordering a ride, completing a payment — without needing app-specific APIs. The company’s agent has held #1 on the AndroidWorld leaderboard since October 2025, the industry benchmark for mobile agent performance.

“We’re making it available on a much bigger scale”

The commercial reality Div described is what makes the thesis credible: this is not a demo. AGI has a direct partnership with Qualcomm to bring its agent stack to Snapdragon-powered devices — smartphones, PCs, and next-generation AI-native devices — “fully private, fully on-device.”

“Running things on the edge, on the device, we’re now making it available on a much bigger scale.”

The deployments he named were concrete and large:

  • Japan’s largest telecom — around 100 million users, roughly the whole country — wants to roll out an agent for the Japanese market: shopping, e-commerce, transportation, all through an agent.
  • Phone and automobile companies, including one of the top three automakers in the world, are exploring AI that powers the car — hands-free voice control of entertainment, navigation, and more.
  • OEM work such as Lenovo’s Qira agent, and deep collaboration with chip players to optimize small models (sub-1B to 8B parameters) for on-device NPUs.

The developer path: AGI.com, an SDK, and an MCP plugin

For builders, Div was direct about how to get on board.

“Go to AGI.com. Anyone can partner with us — we showcase the SDK, and you can download our MCP plugin to set up the platform and run our AI on your laptop, your phone, any device you own. If you want a partnership, just email us and I’ll get back to you.”

The MCP plugin matters: it lets an agent plug into the Model Context Protocol and act on top of the device’s own model, so the intelligence stays with you rather than being rented.

The vision: hands-free, any language, any country

Where this is heading is the most interesting part. Today, people use voice mostly for basic transcription — they can talk, but the assistant cannot yet do things for them hands-free. Div’s goal is the step beyond: an agent you talk to that actually acts, “in any language, in any country.” Privacy and ownership are not features on top of that future — they are the foundation that makes it scale to everyone.

What I took away

  • On-device is the cost fix. Removing the cloud round-trip and the per-token fee is what lets AI reach the next billion people.
  • Ownership is the story. The agent should be yours and run on your hardware or your cloud, not rented from a host that holds your data.
  • It’s already in production — Snapdragon devices, a Japanese telecom at ~100M users, and top-tier automakers — not just a research paper.
  • The developer door is open — AGI.com, an SDK, and an MCP plugin are the on-ramp.

If you are building anything on-device, or you are a partner exploring agentic AI on your own hardware, AGI, Inc. is one of the most concrete companies on that path right now. Start at AGI.com.

Free 30-min Production AI consultation

Book Now