Episode 3 of the AI Agent Bake Off: "Build a GTM Agent for Founders in 72 hours"

#ai #agents #webdev #beginners

Think British Bake Off Show... but with AI Agents instead of Cakes! By the way, I am Abe the host of the show.

We starter with a simple challenge…

The Prompt
Using #Gemini and #ADK build a GTM Agent for Founders that is both MultiModel (beyond text) and MultiAgent (not just an AI wrapper) in 72 hours.

The Setup
And that is what our 3 teams sign up for — 1 developer, 1 googler, an AI Studio API Key, and Google’s Agent Development Kit (ADK).

The Knowledge Base
We also provided them with a Go-To-Market Open Source Repo: https://github.com/goabego/ai-gtm-playbook to get them a head-start

And that's all we gave them (okay also we gave them lunch and plenty of coffee) but the learnings were a bountiful: from context stuffing best practices, to quick A2A deployments (but more on that later)...

What will you see in this episode?

We structured the video in three parts: Day of challenge, Demos, and Judges deliberation.

The Day Challenge

The day of shooting we surprised our teams with a mini 2 hour hackathon to really push their agents to the limit with 6 distinct tests - whoever did the completed the most, wins the challenge - full challenge details here.

Challenge 1: The Launchpad (Deployment)
Challenge 2: The Gauntlet (Load Testing)
Challenge 3: On-the-fly feature (Dynamic Adaptation)
Challenge 4: The Open GTM (MCP Exposure)
Challenge 5: The Ambassador (A2A Exposure)
Challenge 6: The Visionary (Multimodal Input)

All teams where able to accomplish at least 3 in a 2 hour period. Which is a big feat knowing they had no idea what the challenge were going to be and the competitive spirit of the teams. Challenges 1,4,5, and 6 where the was completed in the challenge by the teams.

The Demos

Next you will see the details of the teams building their AI Agents using ADK. You will learn about ADK Web, Gemini, Sub Agent architectures, tactics to manage context and so much more. Here is an example of one of the team's architecture diagram (note all is open source an available below).

Note: All code and reference architecturescan be found here

The Judges

Perhaps my favorite part the Q&A portion between the judges and our teams. In this episode we had Ivan (AI DevRel at Google Cloud), Shubham (AI Product Manager @ Google Cloud), and Annie (AI DevRel at Google Cloud). They asked their questions, shared their thoughts, and selected a winner.

Each team was granted 15 mins of Demo and Q&A back to back. The grading criteria where as follows:

TECHNICAL CRITERIA
Criteria	Assessment	Weightage
Proof of ADK Multi Agent System (MAS)	Showcase and the degree of usage of the ADK Loop, Parallel, and Sequential Agents	25%
Leverage a Gen Media Model	Showcase a the degree of usage of Gen Media model (live api, veo3, image, nano banana, etc)	25%
Handling day of shooting challenge	Showcase and the degree of agent handling the day of challenge	20%

CREATIVE CRITERIA
Criteria	Assessment	Weightage
Impact & Relevance	Level of impact and applicability to real-world problems identified	20%
Presentation & Communication	Clarity and persuasiveness of the solution designed through the pitch video	10%

Note: As well as the winner of the 2 hour challenge was considered in the selection process.

Insights

"Insane" AI agents are not reserved for research labs—it's something you can achieve in a weekend with the right tools like Gemini and the Agent Development Kit (ADK). In this episode, three teams raced to build the ultimate "AI Co-Founder" to rescue startups, and while the 72-hour deadline brought real drama—including surprise hackathons and frantic debugging—the results were accessible and inspiring.

You’ll see that you don't need a complex mesh of confusing code to succeed; judges praised simple, sequential patterns that anyone can learn, and the teams utilized open-source playbooks to fast-track their development. From generating superhero avatars to full validation plans in minutes, this challenge shows that once you break a big problem into small sub-agents, building with AI becomes less about coding magic and more about having fun with creativity.

I think one of our Judges said it best:

The best lesson developers can get from watching this hackathon is that don't build agents for agents sake. Leverage the right structure and frameworks that make sense for your use case. Try to work backgrounds not just forward.

Technical Insights

State Injection > Context Stuffing
- The Insight: "Context Window Management" is the hardest problem in agent engineering.
- What we saw: Team Launchpad relied on massive text files and prompt stuffing (2700% prompt density), which works for prototypes but degrades reliability.
- The Winner's Edge: Team Superpowers (Muhammad & Ayo) used State Injection. Instead of passing the entire chat history to every sub-agent, they extracted specific outputs (e.g., "target_audience_json") and injected only that structured data into the next agent's prompt.
- Takeaway: Don't treat your context window like a trash can. Treat it like a precise function argument.
The "Full-Stack Agent" Reality
- The Insight: An Agent is not just a Python script; it is a UI paradigm.
- What we saw: Every team spent ~50% of their code volume on TypeScript/React (.tsx).
- Why it matters: Agents work asynchronously and often slowly. You cannot just have a loading spinner. You need streaming UIs, intermediate state visualization (like GTM Forge’s dashboard), and "human-in-the-loop" confirmation screens.
- Takeaway: If you are an AI Engineer, you need to learn React (or partner with someone who knows it). The "Headless Agent" is a myth for consumer apps.
Parallelism is the only way to solve Latency
- The Insight: Sequential chains are too slow for real-time users.
- What we saw: Team GTM Forge (Daniel & Luis) reduced a 60-minute linear workflow down to ~15 minutes by running agents in parallel (Map-Reduce pattern).
- Takeaway: Architect your agents to fork. If an agent needs to generate a logo, a blog post, and a video script, those should happen simultaneously, not one after another.
Tools are the new API Standard (MCP)
- The Insight: The Model Context Protocol (MCP) is shifting how agents connect.
- What we saw: All teams had to expose tools via MCP. Team Launchpad used fastmcp for speed, while GTM Forge used the native ADK server for robustness.
- Takeaway: Stop writing custom API wrappers just for your bot. Build MCP Servers. This makes your tool portable—usable not just by your agent, but by any agent (or IDE) that speaks MCP.
Deterministic Guardrails Win
- The Insight: Pure probabilistic reasoning (LLM only) is dangerous for complex tasks.
- What we saw: The winning team (Superpowers) didn't just ask the LLM to "do research." They built a "Shared Preview Server" and local storage mechanisms to save artifacts deterministically.
- Takeaway: "Agentic" doesn't mean "Unstructured." The best agents use rigid schemas (Pydantic/Zod) and deterministic code to handle file saving, API calls, and state transitions, reserving the LLM only for the reasoning parts.

Last but not least, Want a more interactive viewing experience?

We created an even more educational experience with cards, code snippets, and useful links. Try it out here: ai-agent-bakeoff.com you can ask questions and we will answer, we also added the open source content there for easier learning.

Also we open source all the contestants source code and architecture diagrams is available in the first card on the web app.