Google has launched its ADK in Python, Go, TypeScript, Java, and Kotlin.
Google has introduced the ‘Agent Development Kit’ – a major open-source multi-agent framework designed to build reliable, production-ready artificial intelligence (AI) agent systems. Google released the toolkit designed to build, evaluate, debug, and deploy AI agents and multi-agent systems.
While being optimised for Google’s Gemini models, it is model-agnostic and works with other language models as well via LiteLLM.
ADK 2.0 shifts from a rigid, top-down hierarchy to a powerful graph-based execution engine, mixing deterministic code with adaptive AI. It officially supports Python and Go for its 2.0 graph architecture, alongside TypeScript for standard implementation.
In ADK 1.x, agents operated under a strict hierarchical model. ADK 2.0 replaces this with a directed execution graph and reportedly provides dual orchestration:
* Graph-based workflows with explicitly mapped-out node/edge routing with fixed logical gates
* Dynamic workflows which are written in plain, programmatic code
ADK 2.0 also bundles a formal ‘Task API’ for structured agent-to-agent delegation. Instead of confusing back-and-forth prompts, you can implement a defined ‘coordinator-plus-subagents’ pattern that manages multi-turn tasks or single-turn controlled outputs cleanly.
With built-in human-in-the-loop (HITL), human gates are treated as native primitives within the graph runtime. The system can halt execution, await human feedback or approval, and resume from any specified checkpoint.
As version 2.0 completely rewrote the internal execution layer, upgrading requires caution. There are incompatibilities between versions, including storage (using ADK 2.0 code with ADK 1.x databases risks data loss) and event schema changes.
Notably, version 1.0 for Kotlin was released in September for Android and multiplatform ecosystem parity.
