Google ships ADK for Kotlin 1.0, and Android agents can now run on-device

Google announced the 1.0 general availability release of the Agent Development Kit (ADK) for Kotlin on September 9. InfoQ covered the release on September 19, writing that Kotlin has reached feature parity with Google's ADK for Python and Java and gained Android-first extensions for on-device and hybrid AI. The framework is open source on GitHub.

What ships in 1.0

Google says ADK for Kotlin 1.0 brings full parity with the ADK core: hierarchical multi-agent systems where a parent agent delegates to a child, context compaction and multi-turn conversation with history summarization, human-in-the-loop confirmation for sensitive actions, session resumability that lets state be paused, serialized and restored across process restarts, and first-class Java interoperability. The core is built on Kotlin Multiplatform and, in Google's words, "completely agnostic to specific model backends, session providers, or memory systems."

The novel piece is on-device. The Android extensions support LiteRT-LM and ML Kit (currently in beta) for local inference, and Firebase AI Logic for hybrid workflows that run locally when they can and reach the cloud when they must. Agent state can be persisted through Android services: chat sessions in Room, indexed memory in AppSearch, and files in Android storage.

How tools are declared

Tools are described with @Tool and @Param annotations on regular Kotlin functions. Kotlin Symbol Processing (KSP) reads those annotations at compile time and generates the function-call schemas, so there is no runtime reflection to pay for on a mobile device. A tool marked requireConfirmation = true pauses execution and asks the user before it runs, which the announcement suggests for anything with real-world consequence, such as moving money.

Skills, from an SKILL.md file

ADK for Kotlin also loads procedural knowledge from SKILL.md files under src/main/resources/skills/, only when the agent needs them. Google calls the pattern progressive disclosure. The playbook itself never sits inside the model's context by default, which keeps token use predictable even when the library of skills is large.

The most useful thing about the 1.0 for a team that already writes Kotlin on Android is that the whole loop, from typed tool calls through confirmation to persistence, stays in one language and one runtime. An agent that would previously have needed a Python microservice for orchestration and a client SDK for the phone can now be one Kotlin module. On-device inference through LiteRT-LM keeps sensitive prompts and responses on the device, and the cloud is only reached when a larger model is genuinely required.

The tradeoff is that the Android extensions carry most of what makes 1.0 new. Server-side Kotlin gets the multi-agent core and the Vertex AI session, memory and RAG integrations, which is enough to run agent workloads on the JVM without a Python service. The on-device story, the ML Kit path and the Firebase AI Logic bridge, is what the release is built around.

Source

• Primary: Google Developers Blog: Announcing ADK for Kotlin 1.0 (September 9)
• Reporting: InfoQ, September 19
• Repository: github.com/google/adk-kotlin