MINA LABS
Start creating free
Google has released version 1.0 of its Agent Development Kit for Kotlin. The release brings the Kotlin implementation to feature parity with the Python and Java ADK cores, according to the announcement on the Google Developers Blog.
The framework is designed for building AI agents in Kotlin, Android, and other Kotlin Multiplatform targets. It supports multi-agent systems, type-safe function calling, human-in-the-loop workflows, and context compaction. The release also adds Android-focused tools for local models, cloud reasoning, persistent sessions, and semantic memory.
For developers, the main value is the ability to use Kotlin’s existing language and tooling conventions when building agent systems. ADK for Kotlin uses Kotlin Symbol Processing, or KSP, to support zero-reflection function calling. That means developers can define functions for an agent to use while keeping the integration type-safe and aligned with Kotlin’s compile-time approach.
The Kotlin Multiplatform foundation also gives the framework a broader target than Android alone. The same general agent development approach can extend across Kotlin-supported environments, while Android developers get additional integrations designed for mobile applications. This is useful when an agent needs to move between local device capabilities and cloud services rather than relying on a single execution environment.
The Android extensions are the most practical part of the release for mobile teams. Developers can connect local models through LiteRT-LM, use Firebase AI for cloud reasoning, store sessions with Room, and add semantic memory through AppSearch. Those pieces address different parts of an agent product: inference, remote reasoning, conversation state, and retrieval from stored information.
The release also includes orchestration features for more controlled agent behavior. Human-in-the-loop workflows can introduce a person into a process when an action needs review or approval. Context compaction is intended to help manage an agent’s working context as a session grows. Together, these features support workflows that need more structure than a single prompt and response.
For teams deciding where to run an agent, the combination of local and cloud options is important. A mobile application could use an on-device model for some tasks and connect to a cloud model when additional reasoning is needed. The announcement does not define a specific split between those paths, so the right architecture will depend on the application, its model requirements, and how much state it needs to retain.
On Mina Labs, the relevant Gemini Omni Flash five-second generation is available at 17. That gives creators a direct way to test the model experience alongside an agent workflow without treating the Kotlin framework itself as a generation model. The Mina Labs availability is useful for prototyping prompts, agent outputs, and content flows before deciding how those pieces should be connected to a Kotlin or Android application.
We would use ADK for Kotlin to build an Android assistant that can handle a sequence of structured tasks rather than just answer isolated questions. For example, an app could collect a user’s request, call typed functions to work with application data, ask for approval before an important action, and save the session for later use.
We would also test a hybrid setup. LiteRT-LM could handle selected local interactions, while Firebase AI could provide cloud reasoning for tasks that need a larger or more capable model. Room would keep the session history available on the device, and AppSearch could make that history useful as semantic memory. The appeal is not that every task must run in one place. It is that Kotlin developers now have a framework for coordinating these parts in a familiar development environment.