SUMM

Building AI agents offers powerful application possibilities, but productionizing them is challenging due to fragmented frameworks, integration difficulties, and operational complexities
Integrating various tools and frameworks is complex due to lack of standardization
Communication between agents built with different frameworks is not seamless
Managing monitoring, logging, and governance of agents in production is difficult

Google Cloud defines an agent stack with four main components:
- Agent Development Kit (ADK) for building, evaluating, and deploying agents
- Support for protocols (like MCP) to standardize agent-tool and agent-agent interactions
- Vertex AI Engine (Agent Engine) as a managed platform for deploying and managing agents at scale
- Agent-to-Agent Protocol for seamless communication among agents built with different frameworks

Claude and other models can be accessed via Vertex AI’s Model Garden—a centralized hub for discovering and deploying foundational models
Latest Claude models, including Claude 4, are available via Model Garden, ready for testing through the Vertex AI Studio interface
The integration supports both API access and console-based interactions

The demo builds a birthday planner agent using the ADK, organizing party tasks like themes and guest lists
Core ADK concepts: agents, tools, runners (for managing sessions and conversational state), and sessions (for context retention)
Creating an agent requires three files: agent logic (agent.py), environment variables, and an init file
The ADK aligns with software engineering best practices, aiming for easy adoption
Interaction with agents is possible via CLI or a web UI

The system is expanded to include a calendar agent for scheduling, coordinated via an orchestrator agent
MCP protocol enables standard integration of external tools or agents as capabilities within the system
ADK allows easy integration of any MCP-compliant tool with minimal configuration
Multiple agents can be orchestrated to route user requests to the appropriate agent based on context

ADK provides a web interface to run, debug, and view agent interactions, especially useful for multi-agent setups
The UI displays which agents handle specific tasks and tracks the flow of execution

The Vertex AI Agent Engine manages the deployment, scaling, and observability (monitoring, logging, etc.) of agents
It supports agents built in various frameworks (e.g., ADK, LangGraph, LangChain) and various models and tools
The platform provides automatic resource allocation visibility and monitoring for latency, CPU, and memory usage
Deployment requires only a few lines of code, and monitoring is available via the Vertex AI console

To enable cross-framework agent collaboration, Google Cloud introduced the Agent-to-Agent Protocol, an open and enterprise-ready standard
The protocol is based on common industry standards (HTTP, JSON-RPC)
Three main concepts:
- Agent skills (functions/capabilities the agent offers)
- Agent card (digital business card exposing skills to others)
- Agent executor (manages request/response between agents)
This allows agents built on different frameworks to interact and collaborate on complex tasks

The session addressed production challenges: fragmented tools, integration, and operational management
Google Cloud’s solution combines a toolkit (ADK), standardized protocols (MCP, Agent-to-Agent Protocol), and a managed platform (Agent Engine) to streamline agent development and scaling
Attendees are offered QR codes for accessing demo code (ADK samples) and signing up for an upcoming webinar covering advanced integration topics
Presenter encourages questions and offers further support after the session

Building AI agents with Claude in Google Cloud's Vertex AI