Google Releases Kotlin ADK 1.0 for Android and Edge AI Agents

Introduction

Google has released Agent Development Kit (ADK) 1.0 for Kotlin. The production-oriented framework is designed for building AI agents in Kotlin, Android, and JVM or server applications. Google describes the release as functionally aligned with its Python and Java ADK implementations, while adding Android-specific capabilities for on-device and hybrid AI.

For Kotlin teams, the practical implication is that agent logic can be implemented in the same language and engineering environment as the rest of an application. Developers can use idiomatic Kotlin APIs for orchestration, tool execution, persistence, memory, and human-in-the-loop workflows instead of introducing Python solely for the agent layer.

Main capabilities

• Multiplatform execution: The framework is built on Kotlin Multiplatform and can target supported environments ranging from backend services to mobile applications.
• Hierarchical agents: A parent agent can delegate work to child agents, allowing complex workflows to be divided into smaller responsibilities. This capability is useful, but it does not remove the need for clear ownership, permissions, and failure handling.
• Context and session management: Context compaction can summarize conversation history to reduce token usage during long-running interactions. Sessions can be paused, serialized, and restored, which is particularly relevant to mobile processes that may be interrupted.
• Compile-time tool descriptions: Developers can declare tools with annotations such as @Tool and @Param. KSP generates the function description schema during compilation, avoiding runtime reflection and improving type safety and startup behavior.
• Human approval for sensitive tools: A tool can set requireConfirmation to force explicit approval before execution. This is designed for actions with meaningful consequences, such as financial transfers, where a model should not have unrestricted autonomy.
• Progressive skill disclosure: Procedural knowledge stored in SKILL.md files can be loaded when needed. Agents therefore do not have to carry an entire domain manual in every prompt, helping control context size.

Android and hybrid inference

The Kotlin release adds integrations that fit naturally into Android applications. Room can be used to persist conversation sessions, AppSearch can support indexable memories, and Android storage can be accessed directly. For local inference, the framework supports LiteRT-LM, while ML Kit is also available as an experimental option. Cloud and hybrid scenarios can connect to Firebase AI Logic.

This model lets an application choose where an operation should run. A latency-sensitive or relatively simple task may stay on the device, while a task requiring a more capable model can be routed to the cloud. The framework is not restricted to a particular model backend, session provider, or memory system, leaving room for different deployment and infrastructure choices.

Why it matters

The significance of Kotlin ADK 1.0 is broader than adding another language binding. It brings agent development closer to established Android and JVM workflows: tool contracts can be generated at build time, state can be persisted and recovered, sensitive operations can require user approval, and local and remote models can be combined.

There are also important engineering caveats. A production agent should first have a recoverable lifecycle, explicit tool boundaries, and a reliable approval flow. Only after those foundations are in place should teams decide whether a hierarchy of agents is justified. More agents do not automatically produce a safer or more reliable system.

Kotlin ADK is open source. Its combination of multiplatform APIs, Android integrations, and edge-cloud flexibility gives Kotlin developers a direct path to evaluate agent experiences inside mobile and JVM products.

Source: InfoQ Chinese