[ AI Agent ] Google ADK 2.0 Beta Version 研究— 從 Agent 到 Graph：Google ADK 2.0 到底改了什麼？

2026-05-18 · Source: LLM on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, quick

Summary

Google's Agent Development Kit (ADK) 2.0 Beta introduces a significant architectural shift from an "Agent-centric" to a "Graph-centric" design, addressing limitations in controlling complex workflows in ADK 1.x. While ADK 1.x relied on a coordinator LLM to manage agent interactions, ADK 2.0 explicitly defines workflow logic using a `Workflow` object with `edges` in code. This allows for deterministic control over parallel processing, conditional routing, loops, and human-in-the-loop (HITL) interventions, which were challenging to implement reliably in 1.x. The new version treats pure Python functions as first-class nodes within the graph, enhancing testability and auditability. However, ADK 2.0 is currently in Beta, meaning pre-release versions may have breaking changes and it is not compatible with ADK 1.x session data.

Key takeaway

For AI Engineers evaluating agent frameworks for production-grade applications, understand that ADK 2.0 prioritizes explicit workflow control via code-defined graphs over LLM-driven coordination. While it offers enhanced testability and predictability for complex scenarios like parallel tasks and loops, its Beta status and incompatibility with ADK 1.x data mean it is best suited for new prototypes or evaluation in isolated environments, not immediate production deployment or migration of existing systems.

Key insights

ADK 2.0 shifts AI agent orchestration from implicit LLM-driven prompts to explicit, code-defined workflow graphs.

Principles

• Explicit workflow control enhances reliability.
• Code-defined graphs improve testability and auditability.

Method

Define agent workflows as a graph using `Workflow(edges=[...])` to explicitly control parallel execution, routing, and loops, rather than relying on LLM prompt-based coordination.

In practice

• Use ADK 2.0 for new prototypes requiring complex flows.
• Avoid ADK 2.0 for production or existing 1.x data.
• Isolate ADK 2.0 development in a dedicated venv.

Topics

• Google ADK 2.0
• AI Agent Frameworks
• Workflow Orchestration
• Graph-based Workflows
• Prompt Engineering

Code references

• LiuYuWei/adk-two-example

Best for: AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

• Announcing ADK for Java 1.0.0: Building the Future of AI Agents in Java — ADK for Java 1.0.0 enhances AI agent development with advanced tools, control, and interoperability.
• Supercharge your AI agents: The New ADK Integrations Ecosystem — ADK's expanded ecosystem enables AI agents to perform real-world actions across diverse platforms with simple code.
• GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving — GraphFlow uses a unified graph (wGraph) to dynamically manage LLM-agent workflows, improving performance and memory efficiency.
• Building durable Agents with Workflow DevKit & AI SDK - Peter Wielander, Vercel — Workflow DevKit simplifies building durable, observable, and resumable AI agents by abstracting complex orchestration patterns.
• Google Cloud Next ’26 Recap: Welcome to the era of agents — AI agents are transforming work by autonomously performing tasks at scale, supported by new platforms and infrastructure.
• From Workflows to Multi-Agent Systems: How to Choose — Choosing between workflows, single agents, and multi-agent systems depends on task complexity, dynamism, and cost-benefit analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.