Bulding AI Agents with ADK: The Foundation

This codelab, Building AI Agents with ADK: The Foundation, is the first part of a series designed to help you create your own intelligent AI agent using Google's Agent Development Kit (ADK). It guides you through the essential first steps, including setting up your environment and crafting a simple, foundational conversational agent.

The session covers three critical integration patterns: implementing custom Python functions for specific API connectivity, orchestrating specialized sub-agents to utilize built-in capabilities like Google Search, and extending functionality with third-party frameworks like LangChain. Attendees will walk away with a working multi-agent system that demonstrates practical techniques for bridging Large Language Models with real-time data and external services to solve complex tasks.

Transcript

1. Building AI Agents with ADK: The Foundation Fon - Kamolphan Liwprasert Data Team Lead, AIMET.tech Cloud GDE ☁
2. Agents: ADK, A2A & Agent Garden Source: Google I/O ‘25 Developer Keynote ADK Open Source A2A Open Source Agent Garden Open Source Code agents like Google (now production-ready for Python developers) A2A now lets you create seamless and secure agents Sample agents and tools get you started quickly
3. 01 ADK – Agent Development Kit
4. A flexible and modular framework for developing and deploying AI agents. Agent Development Kit (ADK) https://google.github.io/adk-docs Install ADK
5. Higher abstraction makes systems easier to build Low flexibility High flexibility Hard to use Easy to use Solutions: Gemini Enterprise and Conversational Agents Agent-specific framework: Google Agent Development Kit Balance of deterministic control and convenience Low-level orchestration framework: LangChain, GenKit Build your own: Use function calling to build you own agent framework • Smolagent • CrewAI • LangGraph • AG2/Autogen • Agno
6. Google I/O Extended 25 ADK Features https://google.github.io/adk-docs
7. Multi-agent systems greeter story_team researcher writer titler file_writer file_writer_tool payment_agent
8. ADK key concepts https://google.github.io/adk-docs/get-started/about/
9. The Agent Development Kit (ADK) SDK Generative AI Models APIs Functions Databases Agents Tools Profile, goals and instructions short-term long-term Memory Model based Reasoning/Planning Agent Runtime Orchestration (Agent Brain) Key Components • Models: Used to reason over goals, determine the plan and generate a response. • Tools: Fetch data, perform actions or transactions by calling other APIs or services. • Orchestration: Maintain memory and state (including the approach used to plan), tools, data provided and fetched. • Runtime: Execute the system when invoked. End User query response
10. Put it together:
11. blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro
12. Not only Gemini models github.com/BerriAI/litellm
13. 02 Workshop!
14. goo.gle/adk-foundation
15. goo.gle/adk-using-tools
16. adk api_server
17. github.com/google/adk-python
18. 03 Tools
19. ADK Built-in Tool https://google.github.io/adk-docs/tools/built-in-tools/#use-built-in-tools-with-other-tools Google Search Code Execution GKE Code Executor Vertex AI RAG Engine Vertex AI Search BigQuery Spanner Bigtable
20. https://google.github.io/adk-docs/tools/built-in-tools/#use-built-in-tools-with-other-tools
21. Google I/O Extended 25 MCP: Model Context Protocol The Model Context Protocol (MCP) is an open standard designed to standardize how Large Language Models (LLMs) like Gemini and Claude communicate with external applications, data sources, and tools. Think of it as a universal connection mechanism that simplifies how LLMs obtain context, execute actions, and interact with various systems. Image from leapleaper.jp
22. Google I/O Extended 25 https://modelcontextprotocol.io/introduction MCP client-server architecture “How AI can use tools”
23. google.github.io/adk-docs/mcp/#mcp-tools-in-adk googleapis.github.io/genai-toolbox/getting-started/introduction/
24. 03 Multi-Agent System
25. https://google.github.io/adk-docs/agents/
26. Single Agents https://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system
27. Sequential Agents Executes agents in a specific order. For workflows with a fixed order, when tasks have dependencies requiring sequential execution. https://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system
28. Parallel Agents Executes agents simultaneously. For tasks with independent sub-tasks, tasks that prioritize speed, or are resource intensive. https://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system
29. Executes agents repeatedly, until a certain condition is met. For tasks that require iteration, continuous monitoring, cyclical processes or simulated negotiation. Loop Agents https://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system
30. Custom workflow agents Complex workflows Stateful interactions Conditional execution Custom business logic Dynamic agent selection Collaboration or additional perspectives
31. Agent Garden on Vertex AI console.cloud.google.com/vertex-ai/agents/agent-garden
32. ADK Samples https://github.com/google/adk-samples
33. Build with ADK (or any framework), equip with MCP (or any tool), and communicate with A2A, to remote agents, local agents, and humans.
34. Bonus: Agent Starter Pack uvx agent-starter-pack create (new project) uvx agent-starter-pack enhance (existing proj.) pip install agent-starter-pack agent-starter-pack create (new project) agent-starter-pack enhance (existing proj.) หรือ github.com/GoogleCloudPlatform /agent-starter-pack
35. What’s Next? Next Chapter
36. grow.google/ai-professional
37. grow.google/ai-professional
38. Sneak Peek: ChaiyoGCP is back!
41. Google I/O Extended 25 Fon - Kamolphan Liwprasert Data Team Lead, AIMET.tech Cloud GDE ☁󰑆 linkedin.com/in/fonylew