Agentic AI with Google ADK is a hands-on course in building AI agents with Google's Agent Development Kit and Python, from designing agents and tools and managing sessions and memory to multi-agent systems, evaluation and deployment on Google Cloud. It suits Python developers and AI engineers building agents for real work, and you leave with a working multi-agent prototype of your own.
AI agents are moving beyond chatbots that only answer questions towards systems that plan, call tools and carry out multi-step work on people's behalf. Once teams start building them, though, the same questions keep coming up: how to split work across several agents, where to keep conversation context, how to stop an agent from doing things it should not, and how to know an agent behaves correctly before it goes live. Google's Agent Development Kit (ADK) is an open source framework designed to answer these questions in a structured way.
This course takes learners through building AI agents with Google ADK and Python, from a first agent running on a Gemini model to a production-ready multi-agent system. It starts with designing an LlmAgent and writing tools, including function tools, built-in tools, OpenAPI tools and MCP tools, then covers sessions, state and memory and controlling behaviour with callbacks. From there it moves on to dividing work between agents with sub-agents, AgentTool, workflow agents and the graph workflows introduced in ADK 2.0. It closes with evaluation, deployment to Google Cloud and a capstone in which learners build a multi-agent system of their own. (2 days, 6 hours per day, 12 hours in total, Intermediate level.)
Course Details
This course runs for 2 days, 6 hours per day (12 hours in total, 09:00-16:00), as lectures with labs that grow a single scenario throughout, from a first agent to a multi-agent help desk system. Intermediate level. It uses Google ADK for Python with Gemini models through Google AI Studio or Vertex AI. Most labs run on the learner's own machine, and the deployment lab uses the learner's own Google Cloud project. The course focuses on Google ADK itself, covering tools, sessions, callbacks, multi-agent design, ADK 2.0 graph workflows, evaluation and deployment. Learners take home a lab guide, Python sample code for every lab and a capstone starter project.
Section 1: Lab: Getting Started with Google ADK and Your First Agent
What agentic AI is, and how ADK differs from calling an LLM API directly
An overview of ADK 2.0 and its core building blocks: agents, tools, the Runner and sessions
Install ADK for Python, scaffold a project with adk create and configure Gemini through Google AI Studio or Vertex AI
Lab: run a first agent with adk run and test it in the Dev UI with adk web
Section 2: Designing an LlmAgent That Works Precisely
name, description and instruction that help an agent decide correctly, including values from state
Choose a Gemini model, tune generate_content_config and use other models through LiteLLM
Enforce output structure with output_schema and save results to state with output_key
Lab: a product Q&A agent that returns structured data
Section 3: Lab: Function Tools Your Agent Can Call
Turn Python functions into tools with clear type hints and docstrings
Design return values, and use ToolContext to read and write state
Long running function tools for work that waits on a result or an approval
Lab: an agent that checks orders and calculates shipping from sample data
Section 4: Built-in Tools, OpenAPI and MCP Tools
Built-in tools such as Google Search and code execution, and their limitations
Generate tools from an OpenAPI spec and connect MCP servers through McpToolset
Handle tool authentication and credentials securely
Lab: an agent that looks up data through an MCP server and calls an internal API
Section 5: Sessions, State, Memory and Artifacts
How sessions and events hold context, and the different SessionService options
Scoping state with the user:, app: and temp: prefixes
MemoryService for recalling past conversations, and artifacts for files
Lab: an assistant agent that remembers user preferences across sessions
Section 6: Lab: Callbacks, Plugins and Guardrails
Callbacks before and after agent, model and tool execution
Filter requests, mask personal data and validate arguments before calling tools with side effects
Plugins for app-wide policies such as logging and rate limiting
Lab: add guardrails and an audit log to the agent from the previous section
Section 7: Designing Multi-Agent Systems
When to split work across agents, and the coordinator and specialist patterns
sub_agents and transfer compared with calling an agent as a tool through AgentTool
An overview of the A2A protocol: exposing an agent to other systems and calling remote agents
Lab: a help desk system with a triage agent and specialist agents
Section 8: Lab: Sequential, Parallel and Loop Workflow Agents
SequentialAgent for work that runs step by step as a pipeline
ParallelAgent for independent work run side by side, and LoopAgent for draft and review cycles
Passing results between steps with output_key and state
Lab: an automated pipeline that researches, drafts and reviews a report
Section 9: Graph Workflows and Human-in-the-Loop in ADK 2.0
The graph workflow model, where agents, tools and functions become nodes
Conditional routing between nodes, and dynamic workflows controlled by code
Pausing for human approval (human-in-the-loop) before critical steps
Lab: a request approval workflow that asks the user to confirm before acting
Section 10: Lab: Evaluation and Observability
Build eval sets from Dev UI conversations and run them with adk eval
Measure both the tool call trajectory and the quality of responses
Read traces of events and tool calls to find out why an agent answered wrongly
Lab: write a test suite for the help desk system and fix it until it passes
Section 11: Serving Through an API and Deploying to Google Cloud
Use the Runner, read streaming events and expose an API with adk api_server
Compare Agent Runtime, Cloud Run and GKE, then deploy with adk deploy
Manage secrets, persistent sessions, access control and logs in production
Lab: deploy the help desk system to Cloud Run and call it from a web page
Section 12: Workshop: Multi-Agent Capstone
Pick a real-world brief such as a sales assistant, an HR assistant or an internal help desk
Design the agents' roles, tools, state and guardrails
Build the system with ADK, including at least one eval set
Present the work, review it together and wrap up with a pre-production checklist
No public rounds are open right now. Join the waiting list and we will contact you first when the next round opens, or ask us on LINE. Or call 02-570-8449 or 088-807-9770
Google ADK (Agent Development Kit) is a framework for building AI agents in Python. Calling an LLM API directly means writing tool calling, context handling and the decision loop yourself, whereas ADK provides ready-made building blocks such as agents, tools, a runner and sessions, plus a Dev UI for testing, adk eval for evaluation and adk deploy for deploying to Google Cloud.
Write an ordinary Python function with type hints on its parameters and a docstring that clearly states what it does and returns, because ADK uses these to tell the model when to call the tool. Add the function to the tools list of an LlmAgent. Return a dictionary that reports status and result, and use ToolContext when the tool needs to read or write state.
All three are workflow agents that run sub-agents in a predetermined pattern. SequentialAgent runs them one after another like a pipeline, ParallelAgent runs independent sub-agents at the same time, and LoopAgent repeats them in cycles, such as drafting and reviewing until a check passes or a round limit is reached. Results pass between steps through output_key and session state.
It suits Python developers building AI agents for real work, AI engineers and backend developers who want a framework for multi-agent systems, and teams already using Google Cloud or Gemini. You should write working Python, including functions, classes and virtual environments, understand how LLMs and prompting work, and use the command line, Git and REST APIs at a basic level.
You will be able to build Gemini-powered agents with tools that reach external systems through function tools, OpenAPI and MCP, manage sessions, state and memory, add guardrails with callbacks, design multi-agent systems and workflows, evaluate agent quality and deploy to Google Cloud. Participants who complete the training receive a Certificate of Completion with its own number, which can be verified online.
Bring a laptop that can run a Python version supported by ADK and VS Code, plus a Google account for Google AI Studio, as most labs run on your own machine. The deployment lab uses your own Google Cloud project. You take home the lab guide, sample Python code for every lab and the capstone prototype project.
AI today is more than a chatbot that answers questions. Agentic AI can plan and carry out work for us, built around two open standards: MCP (Model Context Protocol), which gives AI a single standard way to use tools, files and databases, and A2A…
Traditional automation follows conditions written in advance, so it stalls on a message that does not fit the pattern or a question that needs data from several places. Agentic AI changes this by letting an LLM interpret the request, plan, and choose which…
Organizations have mostly used AI in a question-and-answer style: a person asks, AI answers, and the person then does the follow-up work every time. Today AI has moved beyond an assistant waiting for commands to an AI agent that takes a goal in human language…
