The AI Agent Bake-Off Ep.1 | Can they build an AI Agent in 3 hours?

Key Concepts

• AI Agents
• Agent Development Kit (ADK)
• Gemini (Google's AI model)
• Imagine (Google's AI model)
• Vio (Google's AI model)
• Virtual Try-On
• Fashion Recommendation
• Multi-Agent Systems
• Context Stuffing
• Live API
• MCP Toolkit
• GenAI Media

AI Agent Bake Off Challenge

The core challenge presented to the four teams was to address the "imagination gap" in online fashion retail. This gap refers to the difficulty customers face in visualizing how clothes will look on them, leading to hesitation in purchasing, high return rates, and negative environmental impacts due to discarded clothing. The teams were tasked with building an AI agent that could help customers "see themselves" in the clothes they want, thereby improving the shopping experience and reducing waste.

Tools and Resources

Teams had access to the following tools and resources:

• Google's AI Models: Gemini, Imagine, and Vio.
• Agent Development Kit (ADK): A platform for building and deploying AI agents.
• Googler "Sue Chefs": Google experts to provide guidance and support.
• MCP Toolkit: A toolkit for GenAI media.

Team Strategies and Approaches

Each team adopted a unique strategy to tackle the challenge:

• Team 1 (Surya and Annie): Focused on a mobile app with a voice agent, skin tone analysis, virtual try-on, and community features (like an affiliate marketing system). They aimed to increase customer loyalty and reduce wasteful spending. They used four agents in ADK: intention identification, skin tone analysis, wardrobe analysis, and general fashion suggestion.
• Team 2 (Tigasa and Ivan): Concentrated on a shopping concierge that provides outfit recommendations based on user preferences, wardrobe contents, and current weather conditions. They emphasized a robust agentic flow and used Gemini Codus and Cursor for development. They also used Google image generation to create pants data, as the provided mock data lacked bottom wear.
• Team 3 (Sam and Lakshmi): Envisioned an AI stylist for Zara, focusing on a live, interactive experience with voice input and a "lending agent" to track borrowed clothing items. They leveraged context stuffing with Gemini 2.5 Pro and the ADK web interface to showcase agent activity. They also used the Live API for voice interaction.
• Team 4 (Dylan and Luis): Experimented with the new MCP toolkit for GenAI media, aiming to generate 360-degree views of users wearing different outfits. They built a simple agent using Imagen, Vio, and the AV tool from the MCP toolkit.

Technical Implementation Details

• ADK Usage: All teams utilized the Agent Development Kit (ADK) to build and manage their AI agents. The ADK's modularity allowed them to tackle individual agent components separately.
• Gemini Integration: Gemini was used for various tasks, including prompt generation, context understanding, and outfit recommendations.
• Image Generation: Google's image generation capabilities were employed to create missing clothing items (e.g., pants) and to visualize virtual try-ons.
• Voice Interaction: Some teams (notably Sam and Lakshmi) integrated voice input using the Gemini Live API, enabling a more natural and interactive user experience.
• Multi-Agent Systems: Several teams implemented multi-agent systems, where different agents were responsible for specific tasks (e.g., intention identification, wardrobe analysis, outfit recommendation).
• Context Stuffing: Team 3 (Sam and Lakshmi) heavily relied on context stuffing with Gemini 2.5 Pro to provide the AI with a comprehensive understanding of the user's preferences and wardrobe.

Challenges Faced

• Time Constraints: The 3-hour time limit proved to be a significant challenge for all teams, forcing them to prioritize features and make difficult trade-offs.
• API Issues: Some teams encountered issues with the virtual try-on API, leading to delays and pivots in their development strategy.
• Environment and Bugs: Teams faced challenges related to development environments, bugs, and the complexities of using new tools.
• Internet Connectivity: Intermittent Wi-Fi issues hampered progress for some teams, particularly in the final minutes of the competition.

Judging Criteria

The judges evaluated the teams based on several criteria, including:

• Creativity and uniqueness of the solution
• Technical implementation and execution
• User experience and design
• Relevance to the challenge prompt

Winner and Rationale

The winning team was Sam and Lakshmi. The judges praised their focus, the unique use of Gemini and ADK, and the seamless integration of the Live API for voice interaction. Their solution felt natural and conversational, creating a "magic moment" where it felt like talking to a real person about clothing styles.

Key Takeaways

• ADK's Power: The Agent Development Kit (ADK) is a powerful tool for building and deploying AI agents, enabling developers to create complex, multi-agent systems relatively quickly.
• Gemini's Versatility: Gemini's capabilities extend beyond simple text generation, encompassing image understanding, voice interaction, and context-aware recommendations.
• Importance of User Experience: A seamless and intuitive user experience is crucial for the adoption of AI-powered applications.
• Challenges of Rapid Development: Building AI solutions in a short timeframe requires careful prioritization, effective teamwork, and the ability to adapt to unexpected challenges.
• The "Magic Moment": The ultimate goal of AI agent development is to create a natural and engaging interaction that feels like talking to a real person.