ADK for Kotlin 1.0: Android Agents, Tools, and Approval
Teams using Kotlin or Android no longer need to migrate everything to Python to start building AI Agents. Google's ADK for Kotlin 1.0 combines multi-agent management, tools, session memory, and user approval in a way that is more compatible with the Kotlin ecosystem.

ADK for Kotlin is a toolkit for creating agents using the JVM family of languages.
Agent Development Kit (ADK) for Kotlin is a framework for building AI agents in Kotlin, supporting Android, server-side work, and Java interoperability. Version 1.0 was announced as matching ADK Core 1.0's main capabilities, with Android and on-device extensions.
The core uses Kotlin Multiplatform and aims to avoid dependence on a single model provider, session system, or memory system, separating agent logic from model location and state storage
A Kotlin example from the README separates the agent name, model, instructions, and search tool.
import com.google.adk.agents.LlmAgent
import com.google.adk.agents.Instructions
import com.google.adk.models.Gemini
import com.google.adk.tools.GoogleSearchTool
val rootAgent = LlmAgent(
name = "search_assistant",
description = "Assistant that can search the web.",
model = Gemini("gemini-3.1-flash-lite-preview"),
instructions = Instructions("You are a helpful assistant. Answer user questions using Google Search when needed."),
tools = listOf(GoogleSearchTool())
)
This example comes from main, whose README identifies version 1.4.0. It does not establish compatibility with version 1.0 or Android without changes. We have not run it; check the documentation for your installed version before use. Source: Source documentation · Apache-2.0
Key capabilities improving practical readiness in version 1.0
- Multi-agent: Arrange agents and assign tasks to those with expertise in each area.
- Context compaction: Summarize history as conversations grow to control context size
- Human-in-the-loop: Pause to request confirmation before consequential actions
- Session resumability: Save and restore state when an app or process is closed and then restarted.
- Type-safe Tools: Uses annotations to generate tool schemas at compile time, reducing runtime reflection
Tools execute actions; skills provide procedural guidance
In Google’s example, tools refers to capabilities an agent can invoke, such as reading metrics, finding deployments, or sending alerts. In contrast, skills provide procedures and rules loaded when needed, such as a playbook for a slow database.
This separation keeps system privileges in verifiable code. While procedural knowledge can be edited and distributed separately from the main agent, a skill is not a sandbox. If a tool has broad permissions, commands in a skill alone cannot guarantee that it will not be misused.
Local execution, persistent state, and approval on Android
ADK for Kotlin supports on-device models through LiteRT-LM and ML Kit, identified as beta in the version 1.0 announcement, and cloud access through Firebase AI Logic. Room and AppSearch integrations support persistence across process restarts. Check extension status for the version you install.
For actions such as transferring money or sending data, show the details and request confirmation. Google’s example applies this to a transaction tool. Enforce approval at the tool or runtime layer rather than relying on the model to remember prompt warnings.
“Production-ready” does not mean our app is immediately ready
Version 1.0 signals a more production-oriented core API and capabilities. Teams still need authentication, authorization, secret storage, retries, observability, rate limits, and recovery from model or downstream-service failures.
Test long context, app closure mid-task, malformed tool output, rejected approvals, and out-of-scope tool requests. Clearly identify data kept locally and sent to cloud.
KrobClick's proposed acceptance test: simulate order reading with approval required before address edits. Use hypothetical data and tools that log requests instead of changing real systems. Test approval, denial, and stopping while approval is pending; resume and inspect status and call counts. Denial must leave data unchanged, and resumption alone must not duplicate calls. This is a plan, not an ADK test result; check the installed version's resume behavior.
Assess ADK for existing Kotlin and JVM workflows
ADK for Kotlin lets Android, Kotlin, and Java teams build agents in familiar languages and structures without moving all orchestration to another stack. Assess provider compatibility, monitoring, testing, and long-term upgrades beyond demos.
A safer start is a workflow that reads information more than it changes systems. Limit tools, require human approval before significant actions, and measure whether quality and time savings justify the added complexity.
Sources
- Google Developers Blog: Announcing ADK for Kotlin 1.0
- Agent Development Kit documentation
- google/adk-kotlin
This article uses Google announcements and documentation as of October 7, 2026. APIs, service integrations, and some extensions' status may change. Review the documentation for your version and test security before production use.