Course Overview
This one-day hackathon challenges participants to build a sophisticated, multi-agent AI system for a fictional newspaper agency. The goal is to automate the article writing process - from initial topic entry to final content moderation - using Google Cloud’s agent and generative AI capabilities with the Agent Development Kit (ADK). The event focuses on accelerating the journey from prototype to production by relying on managed, scalable services for agent deployment, tool connectivity, retrieval (RAG), and governance.
Prerequisites
- General Programming Skills: Proficiency in Python is recommended, as many AI frameworks and tools are Python-based.
- Basic Understanding of Google Cloud: Familiarity with projects, IAM roles/service accounts, and Cloud Run or Cloud Functions will be helpful.
- Fundamental AI Concepts: A basic grasp of large language models (LLMs), prompting, and evaluation is beneficial.
Course Objectives
By the end of the hack, participants will be able to:
- Build, debug, and run agents locally using Google's Agent Development Kit (ADK) developer experience.
- Deploy AI agents at scale on Agent Engine, Google Cloud's managed agent runtime.
- Build and deploy a custom MCP Server on Cloud Run and connect it to an ADK agent to interact with external systems.
- Design and orchestrate multi-agent systems where specialized agents collaborate using ADK's sequential, parallel, and loop agent patterns.
- Leverage Retrieval-Augmented Generation (RAG) using AI RAG Engine and/or Agent Search to ground agents in enterprise data and reduce hallucinations.
- Deploy production-ready agents with Agent Sessions and/or Agent Platform Memory Bank for stateful conversations and implement safety guardrails using Agent Gateway and Gemini built-in safety settings.
Outline: Agentic AI on Google Cloud Platform - 1-day Hackathon (AAI-HACK-G1)
The challenges are interconnected and built upon one another, guiding participants through creating a complex, multi-agent AI system end-to-end. The final product will be a prototype article-writing automation solution that can be deployed on Google Cloud.
Challenge 1: The First Agent
Participants will start by using the Agent Development Kit (ADK) to set up their agent project, run it locally using the ADK developer experience (CLI + web UI), and define a first editorial agent. Teams will then make the agent cloud-ready by packaging it for deployment on Google Cloud - the primary deployment target is Agent Engine, Google Cloud’s managed agent runtime. This first agent becomes the central coordinator for the later multi-agent system.
Challenge 2: Bridging to the Real World
This challenge focuses on giving the agent the ability to interact with external systems. Participants will build and deploy a custom MCP Server on Google Cloud Run, exposing tools such as a web crawler or an article database. Once deployed, they will connect the MCP Server to their ADK agent, enabling it to gather and store relevant sources as the foundation for article writing.
Challenge 3: Multi-Agent Collaboration
Participants will expand their system by creating specialized sub-agents—such as a “Writer Agent” and an “Art Agent” and wiring them into an ADK multi-agent architecture. The challenge culminates in configuring the main orchestrator agent to delegate tasks to these specialized agents and merge results into a single publishing flow. Additionally, teams will explore the possibilities of integrating Retrieval-Augmented Generation (RAG) by grounding the agents on an internal knowledge source using Agent Search and RAG Engine.
Challenge 4: From Prototype to Production
In this final challenge, teams take their multi-agent system to production on Google Cloud. Participants will deploy the full solution to Agent Engine and integrate it with Agent Sessions and/or Agent Platform Memory Bank, enabling persistent, stateful conversations where the agent remembers context across calls. Teams will then harden the deployment by implementing safety guardrails using Agent Gateway for prompt and response filtering alongside Gemini built-in safety settings for content moderation. The challenge concludes with defining a minimal release checklist that demonstrates a credible path from prototype to production.