I Thought AI Agents Were Just Prompts. This Course Proved Me Wrong.

Srinithya Bollena — Sun, 14 Dec 2025 07:08:53 +0000

A zero-to-hero journey through Google & Kaggle’s AI Agents Intensive Course (2025)

I’ll admit it.

On Day 1 of the AI Agents Intensive Course by Google and Kaggle, I thought an AI agent was just:

“A long prompt with confidence.”

Five days later, I was building agents with memory, tools, evaluation, and even other agents coordinating with each other.

Somewhere between "this is easy" and "why is my agent doing that?", my understanding completely changed 😄

This is my zero-to-hero journey.

Day 1: The Day My Prompt Ego Died

Before this course, I believed prompts could solve everything.

Day 1 politely (and then aggressively) proved me wrong.

I learned that an AI agent is not a single response. It’s a system with:

Goals
Memory
State
Tools

Using Gemini and ADK, I built my first agent and then a multi-agent system.

That moment felt like:

“Oh… so I wasn’t building AI. I was just talking to it.”

Day 2: When My Agent Became Actually Useful

Day 2 introduced Model Context Protocol (MCP).

Suddenly, my agent could:

Call real tools
Fetch live data
Pause for human approval
Continue without forgetting everything

This was the moment AI stopped being cool and started being useful.

A good agent knows when not to act.

Honestly, same.

Day 3: Memory — The Difference Between Smart and Forgetful

Before Day 3, my agent forgot everything.

Every conversation started like:

“Hi, who are you again?”

After learning context engineering and memory, my agent:

Remembered past decisions
Stayed consistent
Stopped repeating itself

It no longer felt like a chatbot. It felt like:

A junior teammate who actually reads previous messages.

Day 4: Debugging AI Without Losing My Mind

Day 4 answered a painful question:

“Why did my agent do that?”

I learned observability using:

Logs (what happened)
Traces (why it happened)
Metrics (how good it was)

With LLM-as-a-Judge and Human-in-the-Loop (HITL) evaluation, I could finally improve my agent logically.

AI quality is not magic. It’s engineering.

Day 5: From Demo to Production Thinking

Day 5 connected everything.

I learned about:

Agent-to-Agent (A2A) communication
Multi-agent coordination
Deployment with Vertex AI Agent Engine

This made one thing clear:

AI in production is not about bigger models. It’s about systems that scale and behave reliably.

My Capstone: Thinking Like an Agent Engineer

For my capstone project, I didn’t try to build a “genius agent”.

I built a responsible one.

I focused on:

Clear agent roles
Tool usage with MCP
Memory-aware decisions
Evaluation with logs and metrics

The biggest shift?

I stopped asking:

“How do I improve this prompt?”

And started asking:

“What is this agent responsible for?”

Final Thought

I started this course with zero agent experience.****

I finished it designing agent systems, not just prompts.

And most importantly:

**“I didn’t just learn how to build AI agents.

I learned when they should act, when they should wait,
and when humans should stay in the loop.”**

Huge thanks to Google and Kaggle for creating such a practical and eye-opening course.

I’m never looking at AI the same way again.

Forem: Srinithya Bollena