Google's agent kit reaches 1.0 in Kotlin, with on-device pieces

By Ashlr.AIAI-assisted briefingPublished Sep 20, 2026, 3:00 PM UTC3 min read

Google's Agent Development Kit reached 1.0 for Kotlin, adding Android and JVM targets. InfoQ reviewed it on September 20; Google's own post is dated September 9, and one on-device path is still beta.

What changed

Google published its ADK for Kotlin 1.0 announcement on the Google Developers Blog on September 9, 2026; InfoQ's technical write-up on September 20 brought it to a wider server-side and Android audience. Google's post says the Kotlin release reaches feature parity with ADK Core 1.0 and adds Android-specific extensions on top of a Kotlin Multiplatform core.

Per Google's announcement, 1.0 covers hierarchical multi-agent systems, context management, human-in-the-loop workflows, session persistence, Java interoperability and Vertex AI integration. InfoQ reports that tool schemas are generated at compile time through KSP annotations rather than by runtime reflection, and that sensitive operations can be gated behind an explicit confirmation parameter. Those are build-time and control-flow choices, not model capabilities.

Browse all stories

AI-generated conceptual editorial illustration. Not a documentary photograph.

The briefingAgentic AI

Google launches Gemini 3.8 Live voice models

Sep 18, 20263 min read

Google’s new voice models can keep a conversation moving during background tasks. The useful test is whether dialogue, tool outcomes and access requirements line up.

The on-device story is not one capability. Google's post lists LiteRT-LM and ML Kit for local inference, and labels ML Kit as beta; hybrid cloud work runs through Firebase AI Logic, which the post ties to Gemini 3.8 Flash. Treat each path as a separate availability question rather than assuming your agent runs locally from end to end.

What the announcement does not establish is how any of this behaves in your application. Google's title uses the phrase production-ready; that is the vendor's wording about API stability, not an outside test result. Google's own Android sample carries a note saying it is for demonstration only and is not designed to meet any compliance requirements.

What it means for your work

For builders

The decision this forces is whether agent orchestration still belongs in a Python service that your mobile or JVM app calls over the network. Moving it into Kotlin removes a hop and a deployment surface, and the compile-time schema generation InfoQ describes means a malformed tool definition fails during your build rather than mid-conversation. That is a real debugging difference, worth trying on one small agent. Before committing, test the parts Android will actually punish: session state surviving process death, what happens when the on-device path is unavailable on a given handset, and whether your model backend stays swappable. InfoQ quotes advice to begin with a single resumable agent and explicit tool confirmation before building a hierarchy.

For business

The eligibility question here is data handling, not licensing. An agent that infers on the device keeps prompt content off a vendor's servers, which can change a privacy review — but only for the paths that genuinely run locally, and Google labels the ML Kit route beta. Anything sent through Firebase AI Logic or Vertex AI leaves the device and carries the usual processing terms and metered cost. Ask your team to state, per feature, which path it uses before anyone repeats a claim about local processing in a customer conversation. Also decide who owns the upgrade cadence: a 1.0 label sets an expectation about API stability, not about how often you will be re-testing agent behaviour against new model versions.

Put it to work

Pick one agent feature and write down, tool by tool, which calls could complete with no network at all. Record the inference path each one uses, whether that component is stable or beta in Google's own listing, and what the feature does when the local path is missing. The unanswered question is not whether Kotlin works — it is how much of your agent is actually local. That list decides whether an on-device pilot is worth scoping.

Go to the source

Source material behind this briefing. Our practical implications are Ashlr’s interpretation.

Announcing ADK for Kotlin 1.0: Building Production-Ready AI Agents in Kotlin, Android, and Beyond (opens in a new tab)
Google Developers BlogSep 9, 2026
Google Agent Development Kit for Kotlin Reaches Feature Parity with Python, Supports On-Device AI (opens in a new tab)
InfoQSep 20, 2026

This briefing was prepared with AI assistance from the linked sources. The analysis may include interpretation; consult the original sources for complete context. Cover artwork is illustrative.

Editorial approach Suggest a correction

AI-generated conceptual illustration of processing stations with an inspection gate and a separate recovery path.

Field noteAgentic AI

From AI demos to reliable workflows

Sep 18, 20263 min read

A practical way to turn an impressive AI demonstration into a bounded workflow: define its job, test its failure paths, and decide what counts as finished.