We took a close look at AGI, Inc., the San Francisco lab training small models to see a screen and act on it, with no APIs and no per-app integrations. Inside: why Qualcomm and Lenovo are circling, what 97.4% on AndroidWorld actually proves, and the connector assumption the rest of the agent market can't shake.
In early 2025, OpenAI launched Operator, a ChatGPT agent that can use a computer the way a person does: it opens a browser, looks at the screen, and clicks through booking a flight or filling a form. It was a striking demo of frontier-model screen use, and it worked exactly the way the big labs build things: a huge model running in a datacenter, streamed to your screen, billed per call, and offline the moment your connection drops. Your phone's assistant is a brilliant conversationalist with no hands. Ask it to book a table and it opens the app for you, then leaves. That is where its abilities end, and the product executives at Android OEMs, PC makers, and car companies feel this daily: the apps on their devices expose almost no APIs, per-app integrations break every time a developer ships a redesign, and the fallback, routing every tap through a frontier model's API, adds latency, leaks data off-device, and meters per call. The gap between an assistant that understands a request and a device that executes one is the reason AGI, Inc. exists, and its whole business is closing that gap on the device itself, with no APIs and no datacenter in the loop.
The wedge: the pixels are the API
AGI's bet is that the screen is the only universal interface. Its AGI-0 model family is small, 2B to 20B-plus parameters, and trained to look at a screenshot and decide where to tap. No connectors. No per-app SDK. No accessibility tree. The number that carries the argument: 97.4% task success on AndroidWorld, the highest published score as of early October 2025, achieved across 116 tasks in 20 real apps from screenshots alone. On OSWorld's 369 real desktop tasks spanning Ubuntu, Windows, and macOS, AGI posted 76.26% against a human baseline near 72.36%, which is the rare benchmark where the model beats the person.
The second half of the wedge is the runtime, because a small model only matters if it runs fast on phone-class silicon. AGI-RUN, benchmarked on a Samsung Galaxy S25 Ultra against llama.cpp, Alibaba's MNN, Google's LiteRT, MLC-LLM, and mllm, won 58 of 60 cases at a geomean of 3.3×, up to 9.9× on individual cells. Against on-device SDK startups like Cactus, Nexa, RunAnywhere, and Zetic, the gap ran between 9× and 32×. Copying this is hard for unglamorous reasons. The big labs got where they are by scaling parameters up; squeezing frontier-grade screen understanding into a 2B model runs against how the industry trains. The runtime is grinding systems work, NPU scheduling and prefill throughput, that frontier labs skip and most startups will not staff. And coverage compounds: every app that ships with a visible UI is already supported, no partnership required. Rich Miner, Android's co-founder and now an advisor, compresses it to one line: "It's private, fast, and it works with every app."
The ICP they actually win: device makers with a demo problem
The core buyer is the hardware OEM or silicon vendor whose assistant demos well and then does nothing. Qualcomm's VP of AI, Vinesh Sukumar, announced a collaboration to bring AGI's agent stack to Snapdragon for "smartphones, PCs, and AI-native devices." Lenovo went further at MWC, where its Qira assistant, in a proof-of-concept integration, booked restaurants, ordered rides, and navigated checkouts through the apps' actual UIs, live on stage. Lenovo was careful to call it a PoC, not a product, which is exactly how early supply-chain deals look. These buyers license the model and runtime through a sales call rather than a checkout page, because what they are purchasing is a shipped feature, on their own silicon, with no per-call meter.
Two smaller legs are real. Developers can drive a physical phone from MCP clients like Claude Code, Codex, and Cursor at $0.01 per agent step, with $20 in free credits and sessions capped at 200 steps; the pitch to mobile teams is blunt, since emulators get bot-blocked and some bugs only appear on hardware. And the consumer beta holds a segment most AI marketing talks straight past: blind and low-vision users who need a phone they can operate hands-free. That is demand, not charity.
What the category still gets wrong: it keeps buying connectors
Most of the agent ecosystem optimizes for API surface area. Connector marketplaces. Tool-calling frameworks. Integration counts on pricing pages. The assumption underneath is that apps will eventually expose structured interfaces if you ask nicely, and the fallback is paying a frontier lab per call to do the thinking in a datacenter, the Operator approach. That is the wrong number twice. An agent built on an integration inherits that integration's maintenance burden forever, and it fails on the long tail of apps that will never ship an API, which is most of them. And a per-call meter cannot survive inside a device shipping to a hundred million people, or run in a car with no signal. There is an honesty problem in the evals too: plenty of published GUI-agent results lean on accessibility trees and UI metadata that production apps do not reliably expose, while AGI's leaderboard runs come from pixels alone, the only condition that generalizes to every app a user actually has installed.
The takeaway for operators: the action layer is becoming a component of the bill of materials, sold to the people who make silicon and devices, priced per unit shipped rather than per call, and judged on task success at phone latency. If your agent roadmap assumes APIs exist, watch what AGI is doing with Qualcomm and Lenovo. The leaderboards are interesting. The supply-chain signal is the story.