From 'Why?' to Wow: Building a Multi-Agent Storyteller After 5-Day AI Agents Intensive Course with Google

Source: Dev.to

My “Aha!” Moment: AI Agents Are More Than Just Chatbots

Before the 5‑Day AI Agents Intensive, my view of AI agents was largely centered around conversational interfaces—smart chatbots that could answer questions. The course completely shattered that perception.

The key takeaway was the idea of an agent as an orchestrator of specialized tools. It’s not about one giant model doing everything; it’s about a reasoning engine that breaks a complex problem down and delegates tasks to the best “specialist” for the job. This shift from a monolithic to a modular, tool‑centric mindset was my biggest “aha!” moment.

The deep dive into Multi‑Agent Systems (Day 1) and Tools/MCP (Day 2) was a game‑changer. I stopped thinking about building a single, all‑powerful agent and started thinking about creating a team of “worker agents” managed by a “coordinator”.

Before vs. After

• Before: “How can I prompt a model to generate a story, an image, and audio?”
• After: “How can a Coordinator Agent manage three Specialized Agents—a Writer (Gemini), an Illustrator (Flux.1), and a Narrator (OpenAI TTS)—to work in parallel and deliver a result faster and more efficiently?”

This evolution directly inspired my capstone project.

Capstone Project: Curiosity Storybook

For the Agents for Good track, I built Curiosity Storybook, an AI agent that transforms a child’s “Why?” into a magical, multi‑sensory learning experience. Instead of a dry answer, it generates a complete, personalized storybook page with:

• a story,
• an illustration, and
• an audio narration.

GitHub Repository
YouTube video

General Architecture

1. Frontend (UI/UX) – Kid‑friendly interface built with Gradio, hosted on Hugging Face Spaces.
2. Agent Orchestrator – Main agent managed with Blaxel, using Gemini 2.5 Pro for reasoning and content generation.
3. Tools
   • Custom MCP (Model Context Protocol) server exposing tools for specific tasks like narration.
   • Direct calls to heavy‑compute services for long‑running tasks such as image generation.
4. AI Models
   • Google Gemini 2.5 Pro – Generates the main story and the illustration prompt.
   • Flux.1‑schnell – High‑quality image generation.
   • OpenAI TTS – Audio narration.
   • Hyperbolic (Llama 3.3) – Ultra‑fast generation of related questions.

Building the Project

• Implemented a Coordinator/Specialist pattern: a main agent in Blaxel orchestrates three parallel tasks, each handled by a specialized model. Watching the story, image, and audio generate concurrently proved the power of this architecture for user experience.
• Applied Context Engineering (Day 3): created a ConversationContext class that uses compaction (summarizing history) to feed a “Question Suggester” agent (Hyperbolic). This lets the agent suggest relevant follow‑up questions without needing the entire conversation transcript, keeping the system fast and efficient.
• Integrated Observability from the start (Day 4): added logging for all tool calls and tracing via a session_id. When image generation failed once, the logs pinpointed the exact step, demonstrating the immediate value of this pillar.

Key Lessons

• Multi‑Agent Systems are Practical, Not Just Theoretical – The Coordinator/Specialist pattern works in real‑world applications.
• Context Engineering is the Secret Sauce – Summarizing conversation history enables fast, relevant suggestions.
• Observability Isn’t an Afterthought – Early logging and tracing simplify debugging and improve reliability.

The AI Agents Intensive course shifted my mental model from “prompts” to “systems”. “Curiosity Storybook” is the tangible result of that journey, showcasing agents as complex, problem‑solving engines rather than simple chat interfaces.