Why I Finally Started Learning AI Agents
I’ve been reading and researching so much about AI agents over the past few months and have never settled on any specific methods or frameworks. One reason is that things are evolving so quickly. Whatever AI tools we learn today can easily become obsolete or irrelevant tomorrow. That’s why I’ve held back from focusing on implementation and instead spent time reading research papers and studying different frameworks and structures experimented with by other researchers.
Since it’s already November, I decided it was time to focus on developing hands-on technical skills before the end of the year. A few weeks ago, I came across Kaggle’s 5-Day AI Agents Intensive Course with Google. I immediately signed up, and it became my motivation to finally build some working AI agents.
No-Code, Low-Code, and Full-Code: My Learning Journey
I’ve been focusing on two different areas: entry-level no-code agents and more technical tools that require extensive coding. I briefly tested low-code solutions such as n8n and OpenAI’s Agent Builder but found the learning curve still a bit steep. I wasn’t sure if I wanted to commit the time to master those platforms.
Since my last blog post, I’ve started exploring various ways to build AI agents. For people without a technical background, I found that OpenAI’s GPTs is one of the easiest entry points. Customizable GPTs have been around for a while, but I had never subscribed until recently. ChatGPT remains my favorite LLM for text-related tasks such as grammar checking, drafting emails, and brainstorming presentation ideas. Now, I’m adding Customizable GPT to my toolbox.
What I Learned from Custom GPT
First of all, it’s extremely easy to use. ChatGPT helps users write the right instructions or prompts that define their customization. It also includes built-in tools such as web search and image generation. The most exciting part is “actions”, which allow connections to external tools for specific tasks—for example, connecting to a weather API to get real-time forecasts.
For non-technical users, don’t feel intimidated by schemas (usually JSON or YAML files) used to define actions or functions. In the past, it would have taken me days to figure out how to structure a schema for each API. Now, ChatGPT offers ActionsGPT, which simplifies this process. By providing a URL or API documentation, ActionsGPT helps define the schema automatically—saving a lot of time and effort.
Returning to Google’s Ecosystem
Besides Custom GPT, I’ve also been taking the 5-Day AI Agent Intensive Course, which I’ve found incredibly valuable. A few months ago, I was determined to keep my AI-related projects within the Google ecosystem. However, I eventually gave up because Google’s Agent Development Kit (ADK) lacked sufficient documentation at the time.
I’ve also gotten used to vibe coding. I don’t write code from scratch anymore but act as a code reviewer, analyzing and understanding what LLMs generate. (Just a side note: after testing ChatGPT, Gemini, and Claude, I personally feel Claude performs best on technical tasks.)
At the time, Google’s ADK was so new that most LLMs, including Gemini, were not yet trained on it. Even when they understood it, the generated code quickly became outdated because the ecosystem was evolving so fast. But with the release of this course, I feel the ADK has finally matured enough to support structured learning.
What I Learned from the 5-Day AI Agent Intensive
The course did not disappoint. It’s well-structured with a white paper, a NotebookLM podcast based on the paper, and practical notebooks.
Day 1: Introduced how to build an agent using Google’s ADK and create multi-agent systems. Examples included sequential, parallel, iterative refinement, and LLM orchestrator structures.
Day 2: Focused on creating tools for agents, including customized functions, API connections, and MCP servers. A highlight for me was learning about long-running operations, which enable “AI with a human in the loop.” As a strong advocate for human oversight in investment-related tasks, I found this part especially exciting.
Day 3: Covered context engineering and memory management, showing techniques like RAG and summarization, and how to persist in-session data so LLMs can retain context across conversations.
Day 4: Focused on agent quality and evaluation. It explained how to implement observability within a pipeline and measure performance. Evaluating agents is quite different from evaluating traditional models since there are often multiple correct answers.
Day 5: Discussed how to bring agents into production, including agent-to-agent communication protocols and deployment best practices. I’m still deciding
Conclusion: Turning Learning into Practice
This experience has given me both structure and motivation to continue building real, functional AI agents. While tools and frameworks will keep evolving, the fundamentals, context management, orchestration, memory, and evaluation, remain essential for building reliable systems.
I’m planning to apply what I’ve learned to my investment research workflow, exploring how agents can enhance efficiency and decision-making in finance. I also plan to experiment further with no-code and hybrid tools that make AI development accessible to non-technical users.
As for the capstone project, I haven’t decided whether I’ll complete it yet, but it’s definitely tempting. It would be a great way to put these ideas into practice, consolidate my understanding, and test how far I can push agent design in real-world use cases.
This journey has been incredibly rewarding so far, and I’m excited to keep exploring, experimenting, and sharing my progress as I go.