AI

Google ADK for Kotlin 1.0: AI Agents Running On-Device

Google ADK for Kotlin 1.0: AI Agents Running On-Device
Google ADK Kotlin 1.0 on-device AI agents Android 2026
Google Just Made Android Apps Able to Run AI Agents Without Calling the Cloud
NEW PRODUCT: ADK FOR KOTLIN 1.0

What Google's Own Developer Blog Actually Says

Sep 9, 2026
General availability launch date, per Google's own developer blog[1]
140M+
Android devices already running Gemini Nano, the on-device model ADK for Kotlin can now orchestrate[2]
5 languages
ADK now ships in Python, Java, Go, TypeScript and Kotlin, all built by Google[3]
$0
Cost to use; ADK for Kotlin is fully open source on GitHub[1]

Published: September 21, 2026 | Category: Technology, AI | By Mahesh | Source: Google's official developer blog, primary data current as of September 21, 2026

Today's new product is ADK for Kotlin 1.0, which Google Developer Advocate Guillaume Laforge announced on Google's own developer blog on September 9, 2026. The post opens plainly: "today, we're thrilled to announce the 1.0 general availability release of the Agent Development Kit (ADK) for Kotlin!"[1] Depth Grid has covered the enterprise side of the agent-tooling boom in pieces like OpenAI Will Now Rent You the Same Agent Infrastructure That Runs Codex; this release is the mobile-developer equivalent of that same trend, and it is notable specifically because it is not just another cloud API. It is Google's own toolkit for building AI agents that can run entirely on a phone, with no network connection required.

What Actually Shipped, Read From Google's Own Post

Claim: ADK for Kotlin 1.0 reaches full feature parity with Google's existing Python and Java frameworks, rather than shipping as a stripped-down mobile variant. Source: Google's own post states that "with version 1.0, ADK for Kotlin reaches full feature parity with ADK 1.0 Core while delivering a rich suite of Android-first, on-device extensions," listing the specific capabilities carried over: "hierarchical multi-agent systems" that "chain agents and delegate tasks to specialized child agents"; "context compaction & multi-turn conversations" that "manage context by summarizing history to stay within token limits"; "human-in-the-loop (HITL) & confirmation flows" that "pause execution, request user confirmation for sensitive actions, and resume execution"; "long-running & annotation-based tools" that "automatically generate schemas for tools written in Kotlin"; and "session resumability" to "pause, serialize, and restore active agent interactions across user sessions."[1] The post also confirms "first-party Java interoperability," letting existing Java applications "call ADK Kotlin agents directly."[1] Analysis: Parity matters more than it sounds like it should, because it means a team building an agent for a Kotlin-based Android app is no longer working from a second-tier toolkit compared to teams building server-side agents in Python. That closes a real gap that existed as recently as the 0.1.0 preview release, which Google itself described at the time as "our first experimental version of the library."[4] Published: September 9, 2026.

The On-Device Bet, and Why It's Different From a Chatbot

Claim: The headline capability is not a smarter agent but the ability to run agent reasoning entirely on a phone, with sensitive data never leaving the device. Source: Google's own post states developers can "run fast, private on-device agents using LiteRT-LM and ML Kit (beta), orchestrate hybrid cloud workflows via Firebase AI Logic, or persist agent state across process restarts with Room and AppSearch."[1] A companion Google post on the earlier 0.1.0 preview framed the strategic context directly: "the AI ecosystem is experiencing a massive shift toward the edge, since the introduction of Gemini Nano as a model on Android, it has become available on over 140 million devices."[2] Google's own worked example in the 1.0 post shows a financial assistant agent handling a fund transfer request that "requires explicit user approval" before the actual transfer function executes, with account and transaction data processed through Firebase AI and Android's own Room and AppSearch storage rather than an external server.[1] Analysis: This distinction, between an agent that calls a cloud model for every decision and one that can reason locally using LiteRT-LM (Google's on-device inference runtime), is the actual news buried under the version number. An on-device agent handling banking, health or personal data genuinely reduces what leaves a user's phone, which matters both for privacy and for the agent continuing to function without a network connection, a real constraint for financial, health or field-service apps used in low-connectivity conditions. Published: September 9, 2026.

The Safety Mechanism Google Built In From Day One

Claim: Google designed a mandatory confirmation step directly into the framework for any tool call it considers sensitive, rather than leaving that safeguard to individual app developers to build themselves. Source: Google's own code example shows a banking tool explicitly marked with a requireConfirmation parameter set to true, with the accompanying text explaining that "human-in-the-loop workflows help developers prevent unintended or improper tool usage by requiring explicit human confirmation before sensitive actions are executed," a pattern independent technical coverage confirms is "particularly important for tools that can perform high-impact operations, such as initiating a bank" transfer.[1][3] The framework implements this through what Google calls a "synthetic confirmation request emitted by ADK," which pauses agent execution until the user taps to confirm inside the app's own interface, then resumes the exact same session to complete the action.[1] Analysis: Building this pattern into the framework itself, rather than leaving it as documentation advice individual developers might skip, is a meaningfully safer default. An agent framework that makes the unsafe path, executing a sensitive action without confirmation, the harder path to code rather than the easier one is a genuinely useful design choice as autonomous agents get embedded into more consumer-facing apps handling real financial and personal actions. Published: September 9, 2026.

Why Kotlin Specifically, Not Just Another SDK

Claim: Google built ADK for Kotlin around compile-time code generation specifically to avoid a common performance cost of agent frameworks on mobile hardware. Source: Google's own post states the framework "leverages KSP (Kotlin Symbol Processing) to generate function call definitions at compile time, giving you type-safe schemas, support for suspend functions, and zero runtime reflection."[1] Independent technical coverage of the release quotes a working Android engineer's reaction directly: "commenting the announcement on LinkedIn, PiNCAMP Android engineer Arjun Kumar noted that 'handling tool schemas at compile time with KSP keeps startup fast on mobile targets.'"[3] Analysis: This is a genuinely mobile-specific engineering decision rather than a generic port of a server-side framework. Reflection-based schema generation, the more common approach in server frameworks with abundant memory and CPU headroom, is a real performance liability on a phone where app startup time and battery consumption are directly scrutinised by users; choosing compile-time generation instead is the kind of detail that separates a framework genuinely designed for mobile constraints from one merely adapted to run on mobile. Published: September 9, 2026.

How to Actually Try It

Google's own post points developers directly to the ADK for Kotlin GitHub repository and the official getting-started documentation.[1] To add it to an existing project, Google's post gives the exact Gradle dependency, com.google.adk colon google-adk-kotlin-core version 1.0.0, alongside the KSP processor, with optional Android-specific extensions for ML Kit, LiteRT-LM and Firebase available as separate modules.[1] Developers wanting to see the framework's multi-agent and Java interoperability features in context can review Google's own example gallery and Java interop sample directly on GitHub.[1]

Common Questions

Q1. What is ADK for Kotlin 1.0?
ADK for Kotlin 1.0 is Google's production-ready, open-source framework for building AI agents in Kotlin, announced on Google's official developer blog on September 9, 2026, with full feature parity to Google's existing Python and Java agent frameworks plus Android-specific extensions for running agents directly on a phone.

Q2. Can AI agents built with ADK for Kotlin work without an internet connection?
Yes, for supported use cases: ADK for Kotlin includes on-device extensions using LiteRT-LM and ML Kit that let agents run locally on a phone via Gemini Nano, a model already available on more than 140 million Android devices, rather than requiring every request to reach a cloud server.

Q3. How does ADK for Kotlin handle sensitive actions like money transfers?
Google built a mandatory human-in-the-loop confirmation mechanism directly into the framework, letting developers mark specific tools as requiring explicit user approval, which pauses the agent's execution until the person confirms the action inside the app before it actually runs.

Q4. Is ADK for Kotlin free to use?
Yes, ADK for Kotlin is fully open source and available on GitHub at no cost, though developers using cloud-based models through it, such as Gemini via Firebase AI, would still pay standard usage costs for those underlying model calls.

Q5. How is ADK for Kotlin different from just calling the Gemini API directly?
ADK for Kotlin provides a full orchestration framework on top of any model, including multi-agent delegation, automatic tool schema generation, session persistence across app restarts using Android's Room and AppSearch and built-in confirmation workflows, capabilities a developer would otherwise need to build manually on top of a raw model API call.

Sources

  1. Guillaume Laforge, "Announcing ADK for Kotlin 1.0: Building Production-Ready AI Agents in Kotlin, Android, and Beyond," Google Developers Blog, September 9, 2026. developers.googleblog.com
  2. Google Developers Blog, "Announcing ADK for Kotlin and ADK for Android 0.1.0: Building AI Agents on Android and Beyond," May 21, 2026. developers.googleblog.com
  3. InfoQ, "Google Agent Development Kit for Kotlin Reaches Feature Parity with Python, Supports On-Device AI," September 2026. infoq.com
  4. Android Engineers Substack, "Announcing ADK for Kotlin & ADK for Android 0.1.0," May 21, 2026. androidengineers.substack.com

Read More on Depth Grid

Article by Mahesh | Depth Grid

Discussion (0)

Comments are disabled for this post.