Can we build the ultimate AI co-founder in 72 hours with Gemini?

Key Concepts

Agent Development Kit (ADK): Google Cloud’s framework for building AI agents, offering tools for orchestration, context management, and tool integration.
Multi-Agent Systems: Architectures involving multiple agents working together to achieve a complex task, requiring careful orchestration and context sharing.
Agent Orchestration: The process of coordinating and managing the interactions between multiple agents.
Context Management: Maintaining and sharing relevant information between agents to ensure coherent and informed decision-making.
Model Context Protocol (MCP): A protocol for exposing agent tools to models.
Agent-to-Agent (A2A) Protocol: A protocol enabling direct collaboration and communication between agents.
Sequential Agents: Agents that execute tasks in a predefined order, passing results between each other.
Parallel Agents: Agents that execute tasks concurrently, potentially improving performance.
State Management: Storing and retrieving information about the agent's progress and context.
Looping Constructs: Iterative processes within agents, allowing for refinement and improvement of results.

The 72-Hour AI Agent Bake-Off: A Deep Dive

This document details the proceedings and outcomes of a 72-hour AI agent building competition hosted by Google Cloud, focusing on the challenge of creating a “go-to-market” solution for startups lacking traction. Three teams – Daniel & Luis, IO & Muhammad, and Zach & Lakshmi – competed, each adopting distinct approaches using the Agent Development Kit (ADK). The competition culminated in a demo and judging based on functionality, design, and implementation.

Initial Strategies & Technical Approaches

The teams diverged in their strategies from the outset:

Daniel & Luis: Focused on learning the ADK itself, building an agent to teach them how to use the ADK. They prioritized rapid prototyping and understanding the framework’s capabilities.
IO & Muhammad: Emphasized a clean approach to context flow between agents, building a robust system for information sharing. They leveraged the ADK’s features for seamless communication.
Zach & Lakshmi: Implemented a more complex “inception” approach, building agents that call sub-agents, which in turn call sequential agents. This aimed for modularity and scalability.

The Surprise Challenge & Technical Hurdles

Midway through the competition, Abe presented six challenges designed to test the agents’ real-world applicability:

Deployment to a Public Endpoint: Successfully deploying the agent to a live, accessible URL.
MCP Tool Exposure: Exposing agent tools via the Model Context Protocol.
A2A Collaboration: Enabling direct collaboration between agents using the Agent-to-Agent protocol.
Concurrent User Simulation: Handling a simulated load of 1,000 concurrent users.

Teams encountered various technical difficulties. Zach & Lakshmi struggled with a persistent 403 error during load testing, despite seemingly correct implementation. Daniel & Luis faced challenges with multimodal input, requiring enabling it within Flash (the underlying model). IO & Muhammad focused on building the necessary infrastructure for MCP and A2A.

Team-Specific Implementations & Key Features

1. Daniel & Luis – “Superpowers”

Focus: Building a fun, engaging experience demonstrating the ADK’s capabilities.
Key Features: Multimodal input (image to strategy brief), A2A protocol implementation, and a user-facing website (superpowers.work.com) that transforms user photos into superhero avatars.
Technical Details: Leveraged Flash for multimodal input, utilized the ADK’s deployment option to Cloud Run. They built a system that takes a user’s idea and generates assets like logos, ad videos, and a hosted website.
Quote (Luis): “I more think about the product side. So I think there was a really good match of expertise.”

2. IO & Muhammad – “GTM Forge”

Focus: A comprehensive go-to-market strategy builder for startups.
Key Features: MCP server implementation, A2A protocol integration, a multi-agent system with specialized agents for visual identity, research, and ideation.
Technical Details: Built a wrapper for existing tools and exposed them via the MCP. Employed a multi-agent system where one agent specializes in visual assets and communicates with other agents. Utilized state variables for context sharing.
Quote (Muhammad): “We actually got it working, it was pretty smooth sailing from there.”

3. Zach & Lakshmi – “Launchpad”

Focus: An AI-powered go-to-market strategist providing comprehensive analysis and asset generation.
Key Features: MCP and A2A protocol implementation, sequential agent architecture, integration with Google Search for research, and automated asset generation (social posts, validation plans, website code).
Technical Details: Utilized a sequential agent pipeline with agents for product analysis, ICP (Ideal Customer Profile) creation, and go-to-market strategy generation. Employed a shared search agent for consistent research. Used the ADK’s action confirmation function for user control.
Quote (Lakshmi): “Sometimes you can solve the problem even with a single agent with a very simple multi-agent. You don't have to like over complicated things to achieve something that could have been done easily.”

Challenges & Lessons Learned

Context Management: Maintaining context across multiple agents proved challenging, requiring careful state management and data sharing strategies.
Debugging: The teams encountered unexpected errors (like the 403 error) highlighting the importance of thorough debugging and understanding the underlying infrastructure.
ADK Learning Curve: The ADK, while powerful, has a learning curve. Daniel & Luis’s initial focus on learning the framework proved beneficial.
Parallel vs. Sequential Agents: The teams explored both parallel and sequential agent architectures, demonstrating the trade-offs between performance and complexity.
Importance of Documentation: A recurring theme was the need to thoroughly read and understand the ADK documentation.

Judging & Winner Announcement

The judges evaluated the teams based on functionality, design, and implementation. IO & Muhammad (GTM Forge) were declared the winners, receiving $3,000 in Google Cloud credits. The judges praised their comprehensive solution, thoughtful design, and effective use of multi-agent architecture.

Judges’ Feedback Highlights:

IO & Muhammad demonstrated a strong understanding of agent orchestration and context management.
Zach & Lakshmi’s approach was praised for its simplicity and effectiveness.
Daniel & Luis’s project was commended for its engaging user experience and focus on learning the ADK.

Conclusion

The AI Agent Bake-Off demonstrated the potential of the ADK for building sophisticated AI-powered applications. The competition highlighted the importance of careful design, robust context management, and a deep understanding of the underlying framework. The winning team, IO & Muhammad, showcased a compelling solution for a real-world problem, demonstrating the power of multi-agent systems and the ADK’s capabilities. The event underscored the rapidly evolving landscape of AI agent development and the opportunities for innovation within the Google Cloud ecosystem.