Android AppFunctions Is in Beta. The Code Is Still Alpha

The Android Developers account, AppFunctions, recently announced their service is in beta mode, following its presentation at the Android Dev Summit. The corresponding Jetpack library version is 1.0.0-alpha10. The distinction between this announcement and the actual code is the primary takeaway. Upon reviewing the documentation, it is still labeled as an experimental preview, open to change.

The update preceding the summit was insufficient, as the announcement and the code do not align. The developer experience involves four steps: declaring functions, generating an XML schema via the Jetpack annotation processor, indexing the schema by the OS, and finally, an agent reading the metadata, selecting a function, and executing it with provided arguments.

The only annotation required for the function declaration is @AppFunction(isDescribedByKDoc = true). This means the KDoc comment above the function becomes the interface the agent uses when choosing a tool. The KDoc comment's importance is significant, as it directly influences the function's usage. A simple description like "Creates a task" fails to convey essential details about the function's usage, input requirements, and limitations.

Google's own sample demonstrates a more thorough approach, specifying error handling when necessary input is missing. However, the API has limitations; it requires a minimum compile SDK of 36 and a device running Android 16 or later. As of now, most users are on much older versions of Android, making this API inaccessible to them.

While the hype around AppFunctions portrays it as a groundbreaking invention from Google, the reality is that it utilizes an existing concept: Android-MCP, which has been available for some time. Android-MCP allows agents to interact with a phone through ADB and the accessibility API, working on Android 10 and above. Although faster than traditional screen tapping, it still suffers from latency and layout dependency issues.

Despite these challenges, Google has integrated Gemini, an AI assistant, into apps like Samsung Gallery, Calendar, Notes, and Tasks, showcasing potential use cases. However, this integration is currently in private preview and requires an Early Access Program form for access. While apps can still be controlled using traditional UI automation, this opt-out approach contrasts with the opt-in nature of AppFunctions.

For developers, leveraging AppFunctions is currently more of a learning exercise than a production-ready solution. Many Android apps, particularly those targeting consumers, are still on Android 13 and 14, rendering the API unusable. The Early Access Program further restricts access, leaving a significant portion of the user base unable to benefit from this feature.