ADK 0.1.0: Building AI Agents on Android with Kotlin and On-Device LLMs

Key points

• 1

  Hybrid On-Device/Cloud Workflow: ADK 0.1.0 introduces a hybrid architecture where AI agents can run tasks on-device using local models like Gemini Nano (available on 140M+ Android devices) while offloading complex operations to cloud. This solves privacy concerns by keeping sensitive data offline. For example, a travel app could verify booking details using on-device Gemini Nano without sending data to the cloud. Developers implement this by configuring the orchestrator to delegate specific tasks to on-device subagents via the `LlmAgent` class with `subAgents` and `disallowTransferToPeers` flags. This is critical for apps handling personal data where cloud processing would violate privacy regulations.

• 2

  Kotlin Implementation for Android: To use ADK for Android, add `implementation("com.google.adk:google-adk-kotlin-core-android:0.1.0")` to your `build.gradle.kts`. The framework requires Kotlin annotations like `@Tool` and `@Param` to define functions agents can execute. For instance, creating an 'improbability drive' agent involves defining a `LlmAgent` with a local tool that calculates improbability scores using Gemini Nano. This enables developers to build specialized agents without managing low-level LLM interactions. The key is using the `generatedTools()` method to convert annotated functions into agent tools, ensuring seamless integration with on-device models while maintaining cloud fallbacks for complex tasks.

• 3

  Real-World Use Case: Trip Assistant: ADK powers a trip assistant that handles travel issues by combining on-device and cloud capabilities. When a user reports a booking problem, the cloud-based orchestrator delegates verification tasks to an on-device subagent using Gemini Nano to parse local documents. This keeps sensitive data offline while leveraging cloud reasoning for complex decisions. Developers implement this by setting up a root agent that routes queries to subagents with `subAgents = listOf(heartOfGoldAgent)` and using `disallowTransferToPeers` to prevent unintended data transfers. This approach ensures compliance with privacy laws while providing a responsive user experience without cloud latency.

What changed