Google Dropped Its Agent Framework Into Your Android Phone. Here Is What Actually Works.
Who controls the AI agent on your phone? Google is making its answer easier to build: 140 million Android devices can now run AI agents that process sensitive data entirely on-device, without routing it through a server. The framework landed Wednesday: ADK for Kotlin and ADK for Android, version 0.1.0, the first experimental release of Google's open-source Agent Development Kit for mobile and edge devices.
The headline number holds up. Gemini Nano, the on-device model that powers the Android library, is available on over 140 million devices, according to Google's developer blog. The rest of the announcement is thinner than the number suggests.
The Kotlin ADK library and its Android-specific sibling are now live on Maven Central as version 0.1.0 packages. The GitHub repository exists and the code is accessible. Google describes this release as its "first experimental version," with the library currently supporting default agents for ML Kit GenAI APIs and direct connections to Gemini in the Cloud. The project's own README makes the status explicit: Pre-GA features are "available as is" with "limited support."
That is Google's language, not a journalist's gloss.
The announcement centers on a pattern called hybrid orchestration: a cloud-based model acts as the primary reasoning engine, while specific tasks delegate to sub-agents running fully on-device. In the demo Google showed at its I/O conference, that means a travel assistant that keeps the root reasoning in the cloud but verifies booking confirmations locally, parsing documents through an on-device Gemini Nano model that never routes user data to a server.
The privacy claim is real but bounded. The root orchestrator is in the cloud. Only the sub-agent layer runs on-device. For a trip assistant reading locally stored booking confirmations, that distinction may not matter much to most users. For a medical scheduling app, it matters: the cloud never sees the doctor's name or the appointment time, but it sees that you scheduled something, and when, and for how long. Session metadata exposes patterns even when the content stays local.
The control chain is worth tracing. Google provides ADK and the cloud-based root orchestrator. Device makers provide AICore, the on-device runtime that actually executes the model. App developers build on Google's framework. Users generate the data. Each layer controls a different slice of the stack, and none of them fully owns the session. That gap is where the competitive question lives.
The 140 million device figure deserves scrutiny. Gemini Nano availability requires the "plus" variant of the model, which has more demanding memory and compute requirements. Google has characterized the qualifying hardware as Pixel 8 and newer, and Samsung's Galaxy S24 series, but the blog post does not name specific devices or OS version门槛s. The 140 million likely represents devices that can run it under optimal conditions, not the full Android fleet in circulation.
No production deployments are named in the announcement. No enterprise customers, no case studies, no performance benchmarks. The I/O demo is a trip assistant.
Apple has been building on-device AI into iOS through its Neural Engine for several generations. Qualcomm's Snapdragon chips include dedicated AI inference hardware across a wide device range. Samsung's Galaxy S series ships with its own on-device model. Google's ADK is not the only on-device AI game in town. The difference is that Google is attempting a cross-OEM play: ADK abstracts the device-level runtime so developers write once and the framework routes to whatever on-device silicon is present. Apple and Samsung own their own stacks end-to-end, which makes for tighter integration but narrower reach. Whether ADK's abstraction layer actually works across fragmented Android hardware is the open question.
Google describes the announcement as following recent 1.0.0 releases of ADK for Java and Go, and a beta of ADK for Python 2.0. The Kotlin and Android versions are the new additions, both at 0.1.0.
For developers, the practical value proposition is genuine: a structured way to define multi-agent workflows that route between cloud and device depending on task sensitivity, with session state and memory management handled by the framework. For enterprises, the Pre-GA status and the absence of evaluation tooling, observability features, or SLA guarantees means this is not production-ready infrastructure. Google's own documentation acknowledges that.
What Google announced this week is the beginning of the developer tooling layer for 140 million devices. Whether that layer matures into production infrastructure or remains an I/O demo is the question worth watching.