Mastering AI Agents: Lessons from the Google x Kaggle Intensive & My Capstone

Source: Dev.to

Introduction

Before taking the 5‑Day AI Agents Intensive Course with Google and Kaggle, I viewed Large Language Models (LLMs) primarily as knowledge engines—advanced chatbots that could write code or answer questions. By the end of the week, my perspective had completely shifted. I realized that when you give an LLM tools and a reasoning loop, it stops being just a text generator and becomes an Agent: a system that can perceive, reason, and act.

For my capstone project, I applied these learnings to build a Market Sentiment Monitoring and Alert Agent, a system designed to automate the tedious task of tracking financial news and gauging market health. Below is a reflection on the journey, the core concepts that clicked, and how I built the agent.

The “Aha!” Moment: Tools & Reasoning

The most resonant concept from the course was the transition from Zero‑Shot Prompting to ReAct (Reasoning + Acting) loops.

Before:
I would ask an LLM, “What is the sentiment of the market?” It would hallucinate or give a generic answer based on old training data.

After:
I learned to build an agent that thinks:

Thought: "I need to check the latest news for specific tickers."
Action: Calls a Search Tool or News API.
Observation: Reads the headlines.
Reasoning: "These headlines look bearish. I should calculate a sentiment score."
Final Response: Sends an alert.

Giving the AI “hands” (tools) and a “brain” (Gemini 2.0 Flash) allowed me to build something dynamic.

My Capstone: Market Sentiment Monitoring and Alert Agent

What it Does

• Monitors live financial news for a set of ticker symbols.
• Extracts sentiment‑bearing keywords and assigns a numerical sentiment score.
• Sends an alert when the aggregated score crosses a configurable threshold.

Under the Hood

The Brain: Google Gemini (via the Vertex AI or Gemini API).

The Tools: Integrated Python libraries to fetch live data (e.g., yfinance, web‑scraping tools) and search capabilities.

The Logic:

• Bullish keyword detection – identifies positive terms.
• Context awareness – distinguishes between “loss” in profits vs. “loss” of data.

You can check out the full code and write‑up here: Link to my Project.

Challenges & Key Takeaways

• Hallucinations vs. Reality: Grounding is essential. Using Google Search as a tool ensured the agent referenced real URLs instead of fabricating news.
• Context Windows: Managing how much history the agent remembers (memory) is crucial. You don’t want it analyzing news from 2020 when deciding on today’s alert.
• Evaluation: As emphasized on Day 4 of the course, “vibes” aren’t enough. I manually compared the agent’s sentiment scores against my own intuition to fine‑tune the prompts.

Conclusion

The Google & Kaggle Intensive was a sprint, but it laid the foundation for a marathon. Moving forward, I plan to upgrade my Market Sentiment Agent to not just alert but potentially suggest portfolio adjustments (in a paper‑trading environment, of course!).

If you are on the fence about diving into Agentic AI, start now. The gap between “talking to AI” and “having AI do work for you” is closing fast.