Three ways AI agents will call your Android app
The way Google's agent infrastructure matures directly affects how users will interact with your app. Without structured integration, agents will still use your app, but through slow, error-prone screen simulation with uncontrolled data exposure.
Google is pushing Android toward an agent-aware platform. At the recent Dev Summit, the company detailed three integration paths: a formal MCP-based protocol via AppFunctions, an on-device Agentic API built on ML Kit, and a computer-vision fallback that lets agents see and tap the screen. Each path trades control, privacy, and integration effort. Most apps will likely combine the first two.
The MCP route exposes app features as annotated functions, so agents call them directly without parsing UI. The Agentic API runs a full agent loop inside the app on Gemini Nano 4, which is 3x faster than the previous Nano and keeps data on-device. The fallback, Computer Control Framework, works with zero integration but is slow, fragile, and sees all screen content. Google also showed A2UI, where agents stream JSON to render native Compose UI on the fly, currently in alpha.
The practical takeaway: if you want your app to be agent-friendly, start with AppFunctions and consider Agentic API for embedded assistants. Skipping integration leaves users to a slow, data-exposing fallback that Google itself calls transitional.
The three paths form a spectrum from full developer control to zero integration, with the fallback ensuring adoption but sacrificing user experience and privacy.
Google's strategy mirrors the web ecosystem: structured APIs like MCP are the equivalent of clean APIs, while Computer Control is the screen-scraping fallback for apps that haven't adapted.
The privacy advantage of on-device models is the strongest argument for Agentic API in sensitive categories, but the 3x speed gain also makes it viable for real-time interactions.
A2UI could fundamentally change how UI is built, but its alpha status and the fact that it relies on agent-generated layouts means it will take time to prove reliability.