Agentic AI on Google Cloud Platform - 3-day Hackathon (AAI-HACK-G3)

The challenges are interconnected and built upon one another, guiding participants through the process of creating a complex, multi-agent AI system. The final product will be a complete article-writing automation solution deployed on Google Cloud

Challenge 1: Environment Setup and Agent Foundation

In this initial challenge, participants will establish the core environment for their agentic AI system. This involves setting up foundational Google Cloud services - Cloud Storage for document and asset storage and Cloud Firestore as the article database. Participants will then configure the access to the relevant Gemini models and use the Agent Development Kit (ADK) to create and run their first orchestrator agent locally using the ADK developer experience (CLI + web UI). This agent will serve as the central coordinator for the entire multi-agent system built throughout the hackathon.

Challenge 2: Deploying the Agent’s Toolset

This challenge focuses on giving the agent the ability to interact with external systems. Participants will build a custom MCP Server, exposing tools such as a web crawler and an article database connector. Once running locally, they will connect theMCP Server to their ADK agent, enabling it to gather relevant sources as the foundation for article writing.

Challenge 3: Connecting the Agents to the Tools

Building on the previous challenge, participants will deploy their MCP Server to Google Cloud Run, making it accessible as a remote, scalable service. Teams will containerize their MCP Server, deploy it to Cloud Run, and then configure their orchestrator agent to connect to the deployed endpoint — enabling a fully cloud-hosted tool-calling workflow.

Challenge 4: The Writing and Art Agents

Participants will use ADK to create two specialized sub-agents: a Writer Agent for those drafts of article paragraphs based on the sources gathered, and an Art Agent that generates relevant images using Imagen model. For the Art Agent, participants will create and deploy a dedicated MCP Server to handle image generation, ensuring a clean separation of concerns and appropriate security permissions between agents.

Challenge 5: The Orchestrator

This challenge focuses on wiring everything together. Participants will extend their orchestrator agent using ADK's multi-agent patterns to manage the full editorial workflow: taking an initial topic, delegating research to the MCP tools, spinning up the Writer Agent to draft content, invoking the Art Agent for images, and assembling the results into a single publishable article. The orchestrator becomes the central brain of the entire system.

Challenge 6: Deploying to GCP

In this challenge, participants will transition from a tightly coupled local system to a distributed, production-grade architecture using the Agent-to-Agent (A2A) protocol. Each agent - Orchestrator, Writer, and Art - will be deployed as an independent service on Agent Engine, exposing an A2A-compatible endpoint. The agents will communicate with each other via A2A, enabling them to be managed, scaled, and updated independently. This challenge demonstrates how real production agentic systems are structured as networks of autonomous, interoperable services.

Challenge 7: Making the Multi-Agent System Remember

Participants will extend the Orchestrator agent with persistent state and long-term memory, transforming it from a stateless coordinator into a context-aware editorial assistant. By integrating with Agent Engine Sessions, the Orchestrator agent maintains conversation between requests so editors can return to an article in-progress and pick up exactly where they left off, with topic, sources, and draft state all preserved. Teams can then go further with a long-term memory bank backed by Cloud Firestore or Agent Platform Memory Bank, storing editor preferences, past articles, and recurring topics across sessions for personalized, improving output over time.