Core Update Overview

Google has released ADK for Kotlin 1.0, a production-grade framework for building AI agents in Kotlin, Android, and JVM/server applications. This marks the first time Kotlin achieves feature parity with Python and Java ADK versions, while adding Android-specific capabilities for edge and hybrid AI scenarios.

Key facts:

• Release date: September 2026
• Version: 1.0 (first production-ready release)
• Open source: Yes, available on GitHub
• Vendor lock-in: None—architecture is backend-agnostic
• Platform support: Kotlin Multiplatform, spanning servers, mobile, and JVM environments

Technical Architecture and Features

The core innovation is eliminating Python dependency allowing developers to orchestrate tools, memory, persistence, and human-in-the-loop workflows using idiomatic Kotlin.

Its design admits “no restrictions on specific model backends, session providers, or memory systems”—enabling integration with Firebase AI, Vertex AI, or local LiteRT-LM/ML Kit backends for flexible edge-to-cloud deployments.

Compile-time tool processing via KSP stands out. Using @Tool and @Param annotations, Kotlin Symbol Processing generates function specifications at compile time without runtime reflection. Android engineer Arjun Kumar noted this preserves fast app startup times on mobile.

Human-in-the-loop workflows support sensitive operations with explicit user confirmation. Marking a tool with requireConfirmation = true enforces manual approval:

1
2
3
4
5

@Tool(
    name = "transferFunds",
    requireConfirmation = true
)
fun transferFunds(...)

This integrates natively with Android systems: Room for sessions, AppSearch for indexable memory, and direct file system access.

Multi-Agent and Skill Management

Version 1.0 adds support for hierarchical multi-agent systems, where parent agents delegate tasks to child agents. Context compression and multi-turn conversation summarization reduce token consumption by managing histories automatically.

Session management now includes pause, serialization, and resume capabilities—critical for mobile applications with frequent interruptions.

The “Skills” mechanism uses progressive disclosure: programmatic knowledge stored in SKILL.md files is loaded dynamically only when needed, avoiding full manual ingestion. Joske Vermeulen advises starting with a single recoverable agent plus explicit tool confirmation, scaling to multi-agent systems later—reliability depends more on lifecycle robustness than agent count.

Android-Native Integration

As the Android-optimized variant, ADK for Kotlin 1.0 offers:

• LiteRT-LM: First-class local inference support
• ML Kit: Test version for edge inference
• Firebase AI Logic: Hybrid edge-cloud orchestration

Applications can run lightweight reasoning locally, offloading complex tasks to cloud models seamlessly.

The table below compares feature parity across ADK versions:

Adoption Recommendations

• Adopt now if: Your team uses Kotlin/Android and needs locally-running agents with tool-call orchestration, or requires human confirmation for high-impact operations.

• Wait if: You demand production-grade ML Kit edge inference accuracy, or lack Firebase and local inference infrastructure (ML Kit remains in preview).

Final Thoughts

Kotlin ADK represents a shift from Python-dominated agent development toward multi-language parity. By embedding locality into the language’s ecosystem, mobile devices emerge as autonomous AI actors—not just thin clients.