The Concept: Retro Gaming Meets Modern AI

Abdel Sghiouar and Tilde Thurium demonstrate the "Garden of Forgotten Props," a text-based adventure game designed to teach developers how to build agents using the Google Agent Development Kit (ADK). The game serves as a sandbox for testing agent logic, where players must navigate rooms, manage inventory, and solve puzzles. The project was inspired by classic 1990s text adventures like Zork, updated with a competitive leaderboard to encourage developers to optimize their agents.

Building Agents with ADK

The ADK is a model-agnostic framework that simplifies the process of connecting code to Large Language Models (LLMs) like Gemini. It supports multiple programming languages, including Python, Java, Go, and TypeScript.

Key steps in the development workflow include:

  • Environment Setup: Using Google Cloud Shell provides a pre-configured, browser-based IDE (Visual Studio Code) with 16GB of memory, ideal for quick prototyping without local configuration overhead.
  • Dependency Management: The presenters recommend using uv, a high-performance Python package manager written in Rust, to handle dependencies and virtual environments efficiently.
  • Agent Initialization: The ADK create command launches a wizard that configures the agent's model (e.g., Gemini 2.5 Flash) and access method (Vertex AI or AI Studio API keys).
  • Debugging: The ADK web interface provides a local dashboard for interacting with agents in real-time. Using the --reload-agents flag allows developers to iterate on code without restarting the server, significantly speeding up the feedback loop.

The Role of the Agent

In the context of the game, the agent acts as the player. It receives the game's text output as input, processes the environment, and decides on the next command (e.g., look, take, use, move). The ADK handles the orchestration between the game's API and the LLM, allowing the agent to "reason" through the game's logic. The presenters note that the game is specifically designed to reward the use of Google Cloud infrastructure, adding a layer of gamification to the technical implementation.

Key Takeaways

  • Start with Manual Play: Before automating, play the game manually to understand the command structure and the logic required to solve the puzzles.
  • Use Cloud Shell for Prototyping: Leverage Google Cloud Shell for a zero-setup development environment that comes pre-configured with necessary tools.
  • Iterate Fast: Use the ADK web interface with the --reload-agents flag to test changes to your agent's instructions or logic instantly.
  • Model Agnosticism: While ADK is optimized for Gemini, it is model-agnostic, allowing developers to swap models to see which performs best for specific agent tasks.
  • Manage Secrets Securely: Always use environment variables or secure configuration files for API keys; never commit them to version control.