How to create AI agents with Agent Developer Kit (ADK)

Key Concepts:

• Agent Development Kit (ADK): A Python SDK open-sourced by Google for building agents.
• Agents: Programs with a job and tools, going beyond simple LLM calls to perform actions.
• Tools: Python functions or classes that agents use to interact with the world.
• Model Agnostic: ADK supports various models like Gemini, Llama, Claude, and models available in Model Garden.
• Model Context Protocol (MCP): A protocol for standardizing data inputs for LLMs, integrated into ADK.
• Agent Engine: A runtime environment on Vertex AI for deploying agents with enterprise-grade security and scalability.
• Dev UI: A local development user interface provided by ADK for testing and debugging agents.
• Vibe Coding: Using AI code assistants like Gemini Code Assist to generate code.

1. Introduction to Agent Development Kit (ADK)

• The Agent Development Kit (ADK) is a new Python SDK open-sourced by Google.
• It is designed to facilitate the building of agents.
• ADK is model agnostic and open source, offering flexibility in model selection.

2. Defining Agents and Their Functionality

• An agent is defined as more than just an LLM call; it's a program that performs actions using tools.
• Agents can range in complexity, from simple tool calls to sophisticated systems.
• ADK helps structure tools, which can be Python functions or classes inheriting from a base agent class.
• ADK supports integration with first- and third-party tools like Google Maps with minimal code.
• An agent can also function as a tool or subagent within a larger system.

3. Integration with Model Context Protocol (MCP)

• Model Context Protocol (MCP) is used to standardize data inputs for LLMs.
• ADK integrates MCP, allowing agents to be built as MCP clients or servers.

4. Model Agnosticism and Model Selection

• ADK is model agnostic, allowing developers to use various models, including Gemini, Llama, and Claude.
• Users can select models based on their specific use cases.
• While direct use of custom models isn't fully supported yet, it is a goal for future development.
• Anything available in Model Garden is fair game.

5. Deployment Options

• Agent Engine on Vertex AI is a recommended runtime for deploying agents, offering enterprise-grade security and scalability.
• ADK facilitates simple deployment to Agent Engine.
• Agents can also be deployed to other environments like Cloud Run or virtual machines.

6. Observability and Debugging

• Monitoring agents is crucial, especially as they become more complex.
• ADK provides out-of-the-box logging and tracing on Google Cloud, particularly when deployed to Agent Space.
• ADK includes a dev UI for local development, featuring a chat interface with bidirectional streaming support.
• The dev UI displays every tool call, aiding in debugging and understanding agent behavior.

7. Beyond Chatbots: Diverse Applications

• ADK is not limited to building chatbots; it can be used to create agents that respond to events or other types of input, such as IoT devices.

8. Vibe Coding and AI-Assisted Development

• Vibe coding, or using AI code assistants like Gemini Code Assist, can significantly speed up development.
• AI-assisted coding is particularly useful for greenfield projects.
• Gemini Code Assist can also be used for code review on platforms like GitHub.

9. Conclusion

• ADK is a new and promising tool for building agents.
• The speakers encourage users to try ADK and share their projects and ideas.
• The main takeaways are the flexibility, ease of use, and comprehensive features for developing and deploying AI agents.