Develop AI Agents with Google ADK | MCP | A2A - Masterclass

Install the sdk library and ensure version at least 1.16, then use sdk create to scaffold a first agent in Cloud Shell with an env file, __init__.py, and agent.py.

Create a Gemini API key via Google AI Studio or Vertex AI Studio, linked to a project; key compatibility differs for SDK agents, so use the Google AI Studio key.

Set up a Vertex AI backend for your agent by creating a vertex agent, selecting the same model, configuring the project ID and region, and enabling the Vertex AI API.

Define your first agent using the LM agent or agent, name it route agent, and set model and instruction. Understand using Vertex AI backend, description, and prompt to guide summarization.

Learn to run and test your AI agent with the SDK web interface, performing unit tests in multiple ways before deployment and using debugging tools for quick validation.

Test your agent with SDK run to deploy locally and interact via the command line, with bullet points and pipeline summaries aligning with the web interface.

Learn to run your agent locally with the api server, access it via a restful interface using curl, and manage apps, sessions, and user interactions for production deployments.

Explore the ADK web interface, manage sessions and session IDs, switch between conversations, and use microphone, camera, and file attachments to interact with the agent for debugging and testing.

Run your ai agent locally from your desktop using a gemini api key as the backend, with a PyCharm project and a local UI at 127.0.0.1:8000.

Run your agent from a desktop by using Vertex AI as the backend, authenticate with gcloud's application default login, and optionally use an API key for local setups.

Authenticate an agent with a Google service account for Vertex AI, assign Vertex AI user permission, manage the JSON key, and adopt workload identity federation and Secret Manager.

Learn to use non-google models with ADK, authenticate with api keys from Anthropic or OpenAI, select cloud 37 sonnet latest, and connect via Vertex AI or self-hosted endpoints.

Explore the preview agent designer UI, create a root agent with instructions, generate code, add sub agents and tools like Google search and URL context, and test in the UI.

Explore deployment options for AI agents, including Vertex AI Agent Engine, Cloud Run, and GKE, and consider memory bank and project architecture needs.

Deploy your ADK agent in Vertex AI agent engine using the edk deploy command with a history_agent directory and a staging bucket, then monitor deployment via the GCP logs.

Test the deployed agent with curl by creating a session and querying via the stream query url, then review api urls and runtime metrics in the agent engine console.

Test the agent from a Python program via REST URLs in Colab Enterprise. Create and use a session, handle authentication, and optionally enable streaming responses from the LLMs.

Explore Google's agent starter pack to deploy and create base, agent-to-agent, protocol supported, rag and live agents in minutes, guided by documentation and the Google Cloud Platform GitHub resources.

Create your first agent with the starter pack, naming it and selecting a type. Choose python or go, deploy to agent engine or cloud run, and review the project setup.

Test your agent locally using the make install and make playground commands, view the local agent on port 8501, and customize the app directory and agenda file for different agents.

Deploy your agent to the agent engine with the make deploy command, then verify deployment in the US central region, access it via the provided link, and test end-to-end locally.

Create a live real-time voice and video agent using the agent starter pack, deployable on Cloud Run or any Docker or Kubernetes environment, with a custom UI for testing.

Deploy a live agent to Cloud Run with make deploy, verify its URL and public access, and explore the agent starter pack with UI support for audio, video, and text.

Deploy an academic research agent from the agent garden using the agent starter pack in cloud shell, deploying to the cloud agent engine in 5 to 10 minutes.

Discover how tools transform a model into an agent by connecting external systems, enabling real-time data, advanced calculations, and multi-tool configurations with MCP and third-party tools.

Build an AI agent that uses Google Search as a tool to fetch the latest news from the internet and provide a crisp summary, powered by Vertex AI backend.

Learn to use the BigQuery tool bundled with the SDK to query the employee data set. Configure authentication with application default credentials and set read-only access.

Explore the big query tool as an ai agent that maps tables via dataset ids, answers plain-English questions, and executes sql to reveal employees, managers, engineers, salaries, and leaves.

Define a custom tool by turning Python functions into agent tools, document their capabilities, and show how the language model uses add and multiply functions alongside built-in tools.

Use Langchain tool within ADK to integrate external tools such as the Wikipedia tool. Create a wiki agent that queries Wikipedia via the Wikipedia API wrapper and returns summarized answers.

Demonstrate integrating the GRU AI library and crew AI with the ADK by building a BBC News agent that scrapes a site and retrieves latest headlines using the news tool.

Explore multi-agent systems and how multiple agents coordinate to achieve a larger goal. Compare monolith and microservice-like architectures and preview ADK tools for your first multi-agent system.

Build your first multi-agent system by wiring a route coordinator to delegate tasks to sub agents—joke, story, and song—to greet users, offer services, and deliver tailored outputs.

Demonstrates a multi-agent system with an entertainment agent invoking story and joke sub-agents, controlled by a primary agent, to deliver services and showcase modular agent collaboration.

Modularize your multi-agent system by separating subagents into independent, environment-specific components, enabling individual backends, cleaner imports, and easier testing and maintenance.

Explore a classic multi-agent governance issue where control transfers between agents, causing inconsistent capabilities. Learn how workflow agents can centralize control to prevent this and manage flow.

Master how prompts fix control flow in a multi-agent system by passing control back to the entertainment agent after jokes, songs, or stories. Explore how workflow agents address the issue.

Use subagents as tools by converting them into function calls with the agent tool, keeping control with the route or entertainment agent. Decide when to use tools versus services.

Explore workflow agents, special agents with predefined control flow that orchestrate subagents without relying on large language models, including sequential, loop, and parallel types for deterministic execution.

Understand how sequential agents execute sub agents in a fixed order, collecting data before summarizing, as illustrated by reading a website page then generating its summary.

Learn to implement a sequential agent where a news reader fetches internet news via Google search and a summarizer produces a concise summary, in a strict order.

Demonstrate a sequential agent workflow from a newsreader to a summarizer to produce sport updates (cricket, football, basketball, tennis). Showcases Google search integration and output refinement.

Learn how a sequential agent passes outputs to the next, store results with an output key (for example news output), and use session-scoped state to guide the summarizer agent.

Learn how parallel agents run sub agents in parallel to speed up tasks without dependencies, with the parallel agent orchestrating inputs, collecting outputs, and optionally merging results.

Build a parallel marketing agent that runs brand story and hashtag subagents in parallel under a route agent, detailing non lm and youth-focused prompts.

Run parallel agents in a web demo to develop a watch brand's marketing strategy. Extract two outputs in parallel—a brand story and a hashtag—demonstrating speed without dependency.

Explore loop agent architecture as an orchestrator that iterates through a sequence of subagents, refining Python code through a repeatable review and modify cycle until standards are met.

Demonstrates implementing a code writer agent with a loop of code reviewer and refiner agents using loop and sequential agents, including exit loop and max iterations to produce Python program.

See how the loop agent coordinates sequential and parallel agents, including a code writer, reviewer, and refiner, to produce and polish a Python program via a web demo.

Learn how to structure data for AI agents using JSON input and output schemas, including input schema, output schema, and output key for multi-agent sessions.

Explore input schema in action by validating a country input with Pydantic's base model, demonstrating JSON format data and how instructions enforce or relax schema to fetch a country's capital.

Learn how to define input and output schemas for AI agents, using a JSON-based capital output schema to standardize and validate responses.

Configure an LLM for AI agents by adjusting temperature, top P, safety filters, and output token limits in Vertex AI to balance randomness, safety, and cost.

Set the ai agent's config by adding description and instruction, define generate content config with temperature 0.2 and max output token 400, and apply a safety filter.