Google ADK Masterclass: Build Advanced Multi-Agent AI System

Create your first ai agent—the event planner—by defining the Google ADK agent class blueprint, choosing Gemini 2.5 flash, outlining its expert goals and step-by-step process, and enabling Google search tool.

Turn your agent blueprint into a functioning chat with the ADK runtime's runner, session, and run_async flow. Learn how sessions store memory and how the runner orchestrates queries and responses.

Launch the event planner agent by creating a session and Run Event Planner. Prompt it to plan a 30-person meetup in New York on a $5,000 budget with Google search.

Turn Python functions into custom tools for agents, using docstrings and type hints as the user manual, and manage guests with add guest and get guest list, and JSON dumps.

Apply the specialist agent design pattern by evolving the event planner agent into a team of specialists, including a guest management agent, to handle one task with a robust architecture.

Build a specialized guest management agent with the Gemini 2.5 flash model to add guests and retrieve the current guest list.

Discover how the orchestrator pattern delegates tasks to specialist agents, such as an event planner and guest management agent, acting as a project manager.

Build an orchestrator agent that delegates to two specialist tools, the event planner tool and the guest list manager tool, via the agent tool wrapper.

Run a full multi-turn conversation in a multi-agent system, with an orchestrator delegating tasks to the correct specialist, testing memory, routing, and guest management for event planning.

Explore sequential workflows by building an assembly line of agents that execute tasks in order and pass event plan via output key from the event planner to the communications agent.

Build a plan-and-announce workflow by composing a sequential event planner and event communications agent, wrapped as tools and orchestrated to route tasks and generate announcements.

Apply a loop agent with critique and refine pattern to autonomously optimize a budget, using roles of user, accountant, and cost cutter, looping until the budget fits.

Build the budget optimizer workflow by orchestrating proposer, accountant, cost cutter, and plan retrieval agents to iteratively create, critique, and refine a costed event plan within a set budget.

Build parallel workflows with parallel agents to fan out tasks and fan in results, using catering and entertainment scouts that run Google searches in parallel and synthesize a single report.

Build a parallel vendor search workflow by running catering and entertainment scout agents simultaneously, then synthesize results with a vendor coordinator agent.

Build the capstone autonomous event genie, a master orchestrator that handles complex multi-turn planning and simple user commands, guided by the complete blueprint for our event genie.

Assemble an autonomous event planning workflow by building an intake agent, parallel scouts, and a budget-driven optimization loop, culminating in a final, professional report.

Build the master orchestrator to control the autonomous workflow, wrapping the guest list manager as a tool and defining a robust communications output schema for email drafts.

See the autonomous event Genie run a multi-turn conversation, from agent query and in-memory sessions to guest list management, draft emails, and mock email sending, in a multi-agent demonstration.

Move your code from Google Colab to a local project, set up a virtual environment, arrange the structure, install the Google SDK, and run the Event Genie via ADK web.

Use the adk run command to execute your agent non-interactively from the terminal with a direct prompt, yielding a final answer and exiting, ideal for automation and scripting.

Scale autonomous agents with ADK skills using progressive disclosure to keep the context window lean, load l1 metadata upfront, and fetch l2 and l3 on demand to reduce tokens.

Build and route dynamic ADK skills entirely in memory using models.skill, defining L1 metadata, L2 instructions, and L3 resources to enable automatic routing between vendor negotiation and crisis management.

Learn to add professional observability to a budget optimizer by implementing two ADK callbacks, log-agent-entry and log-agent-exit, to log start/end times, invocation IDs, and session state keys.

Master the multi-agent AI workflow by building single agents, forming agent teams, and applying sequential, iterative, and parallel patterns to capstone projects while connecting to real APIs.

Celebrate finishing the Google ADK Masterclass as you scale from a single agent to a multi-agent architecture, and explore Anthropics' cloud agent SDK for building AI agents in Python.