From Prompt to Agent with ADK

Year: 2025 · ▶ Watch on YouTube

Dr. Fran Hinkelmann (Developer Relations Engineering Manager)

Switch language → zh

Segments (5)

  • 00:00:05 · Introduction to Agent Development Kit (ADK) — Dr. Fran Hinkelmann
    • The speaker introduces the newly released Agent Development Kit (ADK) and its three core components: instruction, tools, and a model.
  • 00:00:40 · Model Context Protocol (MCP) and RAG — Dr. Fran Hinkelmann
    • An explanation of the Model Context Protocol (MCP) and how it enables agents to use tools for Retrieval-Augmented Generation (RAG).
  • 00:01:02 · Live Coding: Building an Agent — Dr. Fran Hinkelmann
    • A step-by-step live coding demonstration of how to define an agent, its instructions, and tools using the ADK Python SDK in Cloud Shell Editor.
  • 00:04:40 · Testing the Agent in the Dev UI — Dr. Fran Hinkelmann
    • The speaker tests the newly built agent using the local ADK Dev UI, providing multimodal input (text and an image) to generate a PDF proposal.
  • 05:33 · Call to Action — Dr. Fran Hinkelmann
    • The speaker announces that ADK is public today and encourages the audience to start building with it.

Products Announced (3)

  • 00:00:08 · Agent Development Kit (ADK) (Publicly available)
    • Python SDK for building AI agents. · Defines agents using three components: Instruction, Tools, and Model. · Includes a local development UI for testing.
    • Publicly available as of the presentation date.
  • 00:00:15 · Vertex AI Agent Engine (Mentioned)
    • The underlying engine that powers the Agent Development Kit.
  • 00:03:35 · Model Context Protocol (MCP) Toolbox for Databases (Contributed to open source)
    • Allows agents to retrieve information from external databases. · Enables Retrieval-Augmented Generation (RAG) capabilities. · Used to connect the agent to a private dataset of building codes.
    • Open source.

Competitor Mentions / Comparisons (1)

  • 00:04:08 · vs Llama, Claude — The speaker mentioned that ADK is model-agnostic and could use models like Llama or Claude, before stating she is using Gemini.

Commitments (1)

  • 00:05:10 (Later this year (2025)) — To integrate the ADK Dev UI with Gemini Code Assist and its supported IDEs.

Demos (1)

  • 01:07 ✓ · Building a Construction Proposal Agent with ADK — Dr. Fran Hinkelmann
    • The demo involved writing Python code in Cloud Shell Editor to define an AI agent using ADK. The agent was configured with instructions, tools (including one for RAG via MCP), and the Gemini 2.5 Pro model. The agent was then run in a local dev UI, where it took a text request and a floorplan image as input and successfully generated a professional, multi-page PDF proposal for a kitchen remodel.

Notable Quotes (4)

  • 00:00:16 — Dr. Fran Hinkelmann:

    To build an agent in ADK with Gemini and Vertex AI, we need three things. We need an instruction, we need tools, and we need a model.

  • 00:00:42 — Dr. Fran Hinkelmann:

    My favorite protocol.

  • 00:01:02 — Dr. Fran Hinkelmann:

    ADK is a Python SDK. Let’s write some code.

  • 00:04:30 — Dr. Fran Hinkelmann:

    In essence, building an agent in ADK boils down to just three things: instructions, tools, and a model.

Visual Signals

On-screen (8)

  • 00:00:05 · Speaker's name and title: Dr. Fran Hinkelmann, Developer Relations Engineering Manager, Google Cloud
    • Introduces the speaker and her role.
  • 00:00:07 · Session Title: From >prompt_ to agent with Agent Development Kit
    • Sets the topic for the presentation.
  • 00:00:15 · Diagram of the Agent Development Kit (ADK) architecture.
    • Visually explains the core components (Instruction, Tools, Model) and their interaction with RAG, APIs, and the Vertex AI Agent Engine.
  • 01:08 · Python code in Cloud Shell Editor defining the root_agent.
    • Shows the simplicity of instantiating an agent with its three key parameters: instruction, tools, and model.
  • 02:34 · Python code for the analyze_building_codes tool function.
    • Demonstrates how to define a tool, including its docstring for the model’s understanding and its connection to an external MCP server for RAG.
  • 04:44 · Terminal command adk web being run.
    • Shows how to launch the local development UI that comes with the ADK.
  • 05:19 · A multi-page PDF document titled 'REMODEL PROPOSAL'.
    • The final, successful output of the agent, demonstrating the end-to-end capability of the demo.
  • 05:34 · A slide with a QR code and URL (goo.gle/adk-docs) to try the Agent Development Kit.
    • Provides a direct call to action for developers to access the documentation and start using the product.

Stage (2)

  • 00:00:05 · Dr. Fran Hinkelmann is introduced and is on stage at a podium.
  • 00:05:25 · Audience applauds after seeing the successfully generated PDF proposal.

Visual demos (2)

  • 01:08 · Live coding in Cloud Shell Editor.
    • The main screen displayed a Cloud Shell Editor window. The speaker navigated between different Python files (agent.py, instruction.py, local_building_codes.py) and a .env file, showing how to configure and define the agent, its instructions, and its tools. A picture-in-picture of the speaker was overlaid.
  • 04:43 · ADK Dev UI walkthrough.
    • The screen showed a terminal launching the adk web server, followed by a web browser displaying the ADK Dev UI. The speaker selected the agent, uploaded a text file and a PNG image, typed a prompt, and showed the agent’s response, which included a link to the generated PDF.

Key Topics

AI Agents · Agent Development Kit (ADK) · Gemini 2.5 · Vertex AI · Model Context Protocol (MCP) · Retrieval-Augmented Generation (RAG) · Python SDK · Developer Tools · Large Language Models (LLMs) · Prompt Engineering · Function Calling · Multimodal AI · Live Coding · Agent Architecture

Takeaways

  • Google has launched the Agent Development Kit (ADK), a new Python SDK designed to simplify the creation of powerful AI agents.
  • Building an agent with ADK is structured around three core concepts: a natural language ‘Instruction’ defining the goal, a set of ‘Tools’ for actions, and a ‘Model’ like Gemini to orchestrate them.
  • ADK integrates the Model Context Protocol (MCP) to connect agents to external data sources and tools, enabling advanced capabilities like Retrieval-Augmented Generation (RAG) from private databases.
  • The kit includes a local development UI (adk web) for rapid testing and iteration, which supports multimodal inputs like text and images.
  • The framework is model-agnostic but is optimized for and demonstrated with Gemini 2.5 on Vertex AI, showcasing complex task completion like generating a professional PDF proposal from a simple prompt and image.
  • ADK is publicly available, empowering developers to start building and deploying sophisticated AI agents immediately.
  • Future plans include integrating the ADK’s development experience directly into IDEs via Gemini Code Assist.