Mobile-Use Agent · AGI-0
AI starts driving
your phone's screen directly.
Until now, the Android version of AGI-0 had stayed a private, invite-only beta. It's now expanding to early access. With no special APIs or plugins, it reads the screen the way a human would and taps, scrolls, and types its way through tasks — and it now tops the leading benchmarks, with an on-device partnership with Qualcomm already underway.
From invite-only
to early access
AGI, Inc. (formerly MultiOn)'s mobile-use agent "AGI-0" has expanded to early access, and the company says it now ranks first on the leading mobile and desktop control benchmarks. Details are on AGI, Inc.'s own blog post, "The World's Most Capable Mobile-Use Agent". Rather than relying on special APIs or plugins, it's built to perceive the screen the way a person would and complete tasks by repeating taps, scrolls, text entry, and app switches — and the scope isn't limited to Android; it extends to browser and desktop control as well.
The benchmark used is Google's own "AndroidWorld," which evaluates agents across 116 tasks spanning 20 real apps — Calendar, Notes, Maps, VLC, Messages, Settings, a web browser, and more. AGI-0 is reported to have posted a score there ahead of existing systems. Because it's scored by driving real apps directly rather than a purpose-built test harness, the benchmark tracks closer to actual user-facing difficulty than a score tuned for a synthetic environment would.
What the benchmark
says it reached
Moving processing
on-device with Qualcomm
Alongside cloud-based inference, work is underway to run the agent locally on Snapdragon-powered hardware.
AGI, Inc. has formed a formal partnership with Qualcomm to bring agentic processing to Snapdragon-powered hardware in a private, on-device form. Moving away from round-tripping every action to a cloud API and toward on-device processing would cut latency from the back-and-forth and ease concerns about action data passing through an external server. Because a phone-control agent inevitably touches sensitive actions like password entry and payments, this direction carries real practical weight, not just a technical footnote.
If cloud and on-device versions end up coexisting, a division of labor becomes possible: routine, repetitive actions handled locally, with only tasks that require complex judgment routed to a larger cloud model. That kind of split is the sort of thing a hardware partnership like this one is positioned to actually implement.
Not waiting on an API —
it just looks at the screen and acts.
Who it helps,
and what comes next
Engineers gain another option for automating apps that don't expose an official API — since it works from screen perception, there's no need to wait on the target app's own integration or update. For PMs, it's one more candidate worth evaluating for verifying routine internal mobile workflows, like spot-checking an approval flow. If browser automation alone already covers your needs, this news is mobile-specific and only loosely relevant. If you're considering joining early access, the first useful step is checking whether the actions you want to automate resemble the kind of apps AndroidWorld actually evaluates — that's what keeps expectations calibrated.
Screens change,
and so does room for error
Topping a benchmark reflects performance on a specific task set — it doesn't guarantee the same accuracy across every app and every UI change. Compared to API-based automation, screen-perception agents structurally retain some risk of mis-taps when a layout updates or a dialog's wording changes. For tasks touching payments or personal data, it's prudent to start where you can review the execution log, and keep a human as the final approver on sensitive actions for now. A detailed error-rate breakdown hasn't been published yet, and real-world data from early access will be the next thing worth watching.
Even with the expansion beyond invite-only, this is still a limited early-access stage. Reaching general availability will require verifying things a benchmark score alone can't capture — stable behavior across a wide range of devices and OS versions, and resilience against adversarial UI changes from apps actively trying to resist automation.