Can Google's ADK Replace LangChain and MCP? (with Christina Lin)

The podcast explores how to build systems using AI, focusing on Google's approach and the balance between opportunity and existential threat that AI presents. Christina Lin from Google discusses the company's internal shift towards AI-driven software development, noting that 25% of their code is now AI-generated. She also addresses the evolving nature of search, with Google prioritizing summarized results while maintaining detailed options. The conversation covers Google's Agent Development Kit (ADK), a Python toolkit for building agentic development pipelines, and A2A, a protocol for agents to communicate directly. Lin emphasizes the importance of responsible AI innovation and the potential for AI to revolutionize how we interact with computers through multi-modal search and real-time virtual realities.

Outlines

Part 1: AI's Impact on Google

Part 2: Agentic Systems and Architecture

Part 3: Tools, MCP, and Intent Discovery

Part 4: Development Best Practices

Part 5: Testing, Debugging, and Deployment

Part 6: Memory, Security, and Advanced Protocols

Part 7: Future Outlook and Getting Started

Sign in to continue reading, translating and more.

Continue

Developer Voices

Part 1: AI's Impact on Google

00:00AI's Existential Threat and Massive Opportunity for Google

AI's Existential Threat and Massive Opportunity for Google

01:56Google's Agent Development Kit: A Python Toolkit for Agentic Development

Google's Agent Development Kit: A Python Toolkit for Agentic Development

03:41AI's Impact on Google: Efficiency in Software Development and Search Evolution

AI's Impact on Google: Efficiency in Software Development and Search Evolution

05:22Google's Leadership in AI Compliance and Multi-Modality Search

Google's Leadership in AI Compliance and Multi-Modality Search

07:50The Future of Computer Interaction: Real-Time Virtual Realities and Agent Development Kit

The Future of Computer Interaction: Real-Time Virtual Realities and Agent Development Kit

Part 2: Agentic Systems and Architecture

09:41From Static Pipelines to Complex Workflows: The Need for Agentic Systems

From Static Pipelines to Complex Workflows: The Need for Agentic Systems

11:20Design Choices in Agentic Systems: Parallelism, Sequencing, and Model Selection

Design Choices in Agentic Systems: Parallelism, Sequencing, and Model Selection

12:45Defining Agents: Autonomous Thinking and Action Beyond LLMs

Defining Agents: Autonomous Thinking and Action Beyond LLMs

14:21The Architecture of an Agent: LLM as Brain, Prompts as Directions, Tools as Actions

The Architecture of an Agent: LLM as Brain, Prompts as Directions, Tools as Actions

15:58ADK as a Framework: Opinions on Agent Structure and Mandatory Components

ADK as a Framework: Opinions on Agent Structure and Mandatory Components

Part 3: Tools, MCP, and Intent Discovery

17:28Building a Version Control Agent: Adapting Tools for New Platforms

Building a Version Control Agent: Adapting Tools for New Platforms

19:01Toolboxes and Agent Intelligence: How LLMs Discover and Utilize Tools

Toolboxes and Agent Intelligence: How LLMs Discover and Utilize Tools

20:49Addressing Knowledge Gaps: Defining Tools and Leveraging MCP for Intent Discovery

Addressing Knowledge Gaps: Defining Tools and Leveraging MCP for Intent Discovery

22:31MCP Server: Simplifying Tool Execution and Intent Recognition for LLMs

MCP Server: Simplifying Tool Execution and Intent Recognition for LLMs

24:34Building Agents: Finding and Annotating Functions for LLMs

Building Agents: Finding and Annotating Functions for LLMs

Part 4: Development Best Practices

26:28Flexibility in Model Selection and the Dark Art of System Prompts

Flexibility in Model Selection and the Dark Art of System Prompts

27:27Building a Friday Night Out Agent: Choosing LLMs and Defining Fun

Building a Friday Night Out Agent: Choosing LLMs and Defining Fun

29:21Grounding vs. Search: Citations and Fact-Checking in LLM Responses

Grounding vs. Search: Citations and Fact-Checking in LLM Responses

30:53Pipeline Complexity: Balancing System Prompts and Agent Responsibilities

Pipeline Complexity: Balancing System Prompts and Agent Responsibilities

33:41Agent System Complexity: Microservices and Traditional Programming Principles

Agent System Complexity: Microservices and Traditional Programming Principles

35:14When to Use Agents: Processing Difficult Workflows and Avoiding Vibe Coding

When to Use Agents: Processing Difficult Workflows and Avoiding Vibe Coding

37:40LLMs and Happy Paths: The Challenge of Edge Cases and Exception Handling

LLMs and Happy Paths: The Challenge of Edge Cases and Exception Handling

Part 5: Testing, Debugging, and Deployment

38:15Debugging Agents: ADK Web and Tracing Tools

Debugging Agents: ADK Web and Tracing Tools

39:54Testing Agent Behavior: Tool Trajectories and Response Similarity

Testing Agent Behavior: Tool Trajectories and Response Similarity

41:37Test Case Examples: Version Control and Night Out Agents

Test Case Examples: Version Control and Night Out Agents

42:45Comparing Natural Language Responses: Rouge Algorithm and Similarity Scores

Comparing Natural Language Responses: Rouge Algorithm and Similarity Scores

44:45Testing Options: Single Tests, Evaluation Sets, and Conversational Interactions

Testing Options: Single Tests, Evaluation Sets, and Conversational Interactions

46:22Deployment: Traditional Microservices and LLM Considerations

Deployment: Traditional Microservices and LLM Considerations

Part 6: Memory, Security, and Advanced Protocols

47:23Memory Management: Stateful Agents and Session Storage

Memory Management: Stateful Agents and Session Storage

49:21Safety Guardrails: Policies and Context for Customer Support Agents

Safety Guardrails: Policies and Context for Customer Support Agents

51:26Security Measures: Model Armor and MCP Server Authentication

Security Measures: Model Armor and MCP Server Authentication

53:07Building MCP Tools with ADK: Integration and Tool Definition

Building MCP Tools with ADK: Integration and Tool Definition

54:10RAG Integration: Data Retrieval as Part of the Tooling

RAG Integration: Data Retrieval as Part of the Tooling

55:15MCP vs. A2A: Agent-to-Tooling vs. Agent-to-Agent Communication

MCP vs. A2A: Agent-to-Tooling vs. Agent-to-Agent Communication

57:05Beyond Text: Agent Cards and Discovery Mechanisms in A2A Protocol

Beyond Text: Agent Cards and Discovery Mechanisms in A2A Protocol

Part 7: Future Outlook and Getting Started

58:12Distributing Sub-Agents: Local Instances and Remote Agents