DEV Community

naoki_JPN
naoki_JPN

Posted on • Originally published at zenn.dev

Building Production AI Agents with Google Cloud ADK + Claude [30-min Workshop]

#googlecloud #claude #agents #mcp

Note: This article summarizes the following X post video (approx. 30 min) in English.
Speaker: Ivan Nardini (Google Cloud Developer Relations Engineer, AI/ML) / Recorded at an Anthropic-hosted event.
Original YouTube: Building AI agents with Claude in Google Cloud's Vertex AI | Code w/ Claude

Introduction

You've built an AI agent — but can't ship it to production. That's the wall Ivan Nardini (Google Cloud) dismantles in this 30-minute workshop.

Using ADK, MCP, Vertex AI Agent Engine, and A2A Protocol, he walks through building and deploying a multi-agent system powered by Claude — end to end.

Why AI Agents Are Hard to Productionize

Prototypes are easy. Production is hard. Three root causes:

Challenge Details
Fragmented landscape Too many frameworks — unclear what to choose
Hard to integrate Cross-framework agent communication is complex
Lack of ops & governance Monitoring, logging, and scaling must all be hand-rolled

Google Cloud's Agentic Stack is designed to solve all three.

The Google Cloud Agentic Stack

Four layers, each targeting one of the above challenges:

Layer Role
Agent Development Kit (ADK) Open-source, code-first agent development framework
Model Context Protocol (MCP) Open protocol standardizing how apps provide context to LLMs
Vertex AI Agent Engine Managed platform for deploying and scaling agents in production
Agent2Agent (A2A) Protocol Open standard enabling cross-framework agent collaboration

Demo 1: Build Your First Agent with 3 Files

Using a birthday planner agent as the example:

from google.adk.agents import LlmAgent
from google.adk.models.anthropic_llm import Claude
from google.adk.models.registry import LLMRegistry

root_agent = LlmAgent(
    name="birthday_planner",
    model="claude-3-7-sonnet@20250219",
    description="An agent that helps plan birthday parties",
    instruction="Handle guest lists, venue suggestions, and scheduling..."
)
Enter fullscreen mode Exit fullscreen mode

Just three files: agent.py, .env, requirements.txt. One command to run:

adk run birthday_planner    # CLI interaction
adk web                     # Browser UI + debug view
Enter fullscreen mode Exit fullscreen mode

ADK supports LlmAgent, SequentialAgent, and other patterns — compatible with Claude, Gemini, and more.

Demo 2: Go Multi-Agent with MCP

To extend the birthday planner to also schedule calendar events, you add two more agents and an orchestrator:

  • BirthdayPlannerAgent — party suggestions
  • CalendarServiceAgent — calendar operations via MCP server
  • EventOrganizerAgent — routes requests to the right agent

Connecting an MCP server is two lines:

mcp_tools, exit_stack = await MCPToolset.from_server(
    connection_params=SseServerParams(url=MCP_CALENDAR_SERVER_URL)
)

agent = LlmAgent(
    name="CalendarServiceAgent",
    model="claude-3-7-sonnet@20250219",
    tools=mcp_tools,
    ...
)
Enter fullscreen mode Exit fullscreen mode

Any existing MCP server can be plugged in as a tool. The orchestrator auto-routes requests based on agent descriptions.

Demo 3: Deploy to Vertex AI Agent Engine 

agent_engines.create(
    agent=root_agent,
    requirements=["google-cloud-aiplatform[adk]"]
)
Enter fullscreen mode Exit fullscreen mode

What you get automatically after deploy:

  • Observability via Cloud Trace / Logging / Monitoring
  • Session management (persistent conversation history)
  • Integration with Vertex AI Evaluation Service for continuous improvement

Works with LangGraph, LangChain, LlamaIndex, and CrewAI too — not just ADK.

Bonus: A2A Protocol — Cross-Framework Agent Communication

When you need a LangChain agent and an ADK agent to collaborate, you need a shared language: Agent2Agent (A2A) Protocol.

Two core concepts:

  • Agent Card: A digital business card for the agent — lets other agents discover what it can do
  • Agent Skills: Describes the agent's specific capabilities and API

Built on HTTP / JSON-RPC, enterprise-ready security included.

Summary

Takeaway Detail
ADK: 3 files, 1 command Fastest path to a working agent
MCP: 2 lines Plug in any existing MCP server as a tool
Agent Engine: zero-ops deploy Observability, scaling, sessions — all managed
A2A: break the framework wall Claude, Gemini, LangChain, CrewAI can coexist

ADK + MCP + Agent Engine + A2A gives you a complete stack from local dev to production scale.

Top comments (0)