agent-building & vibe-coding ?
During my Japanese learning journey, I’ve grown increasingly tired of constantly switching between different apps and browser windows just to keep my study workflow going.
By the way, I think part of the reason this learning process has felt so messy is that language learning itself is no longer linear or traditional. I don’t study with a single physical textbook anymore; most of my learning materials live online, scattered across different platforms and formats. I also don’t rely on just one source, videos, digital textbooks, notes, and examples all come from different places.
On top of that, AI has made it possible to deeply personalize how we learn. This is incredibly powerful, but it also makes the learning process harder to manage. Each study session can easily become overwhelming and frustrating, not because I lack motivation, but because I’m constantly trying to hold together too many moving parts.
Fortunately, with the AI agent I built two days ago, I’ve now simplified this workflow dramatically. All I need to do now is manually paste the set of sentences I want to internalize, and the rest of the process is handled automatically by the agent, which honestly feels incredible.
That said, I’m very aware that I still have a lot to learn. And I tend to learn best through comparison, especially by deliberately placing side by side concepts that easily confuse me or feel hard to distinguish at first. For example, “vibe coding” and “ai agents” sounded almost interchangeable to me at first.
After building my own AI agent hands-on, I also started watching some very introductory videos by Andrew Ng about building products without writing code. Even though I already have some coding experience and find the material quite simple, I still catch myself struggling to clearly articulate the difference between these two ideas.
To clarify my own understanding, I asked ChatGPT to help me distinguish between them. More specifically, I wanted to compare what I had just done — building an AI agent using n8n — with the idea of turning that same n8n automated workflow into a website or an app. I wanted to understand what additional work would be required in that case, and whether it would even make sense to do so, or if it would be better to stick with n8n since my use case is fundamentally about workflow automation.
My original workflow overview
I follow an online Japanese course and read the textbook as a PDF on my Mac; as I study, I copy a set of important example sentences that I want to both understand deeply and eventually produce automatically in speech, and paste this same set of sentences into a Notion page that already contains two instruction types—one asking ChatGPT to analyze and explain the sentences as long-term knowledge, and the other asking ChatGPT to process them for listening practice.
I then copy the entire Notion page into ChatGPT, paste the explanation-focused output back into Notion as a knowledge asset, and paste the listening-focused output into a text-to-speech app for repeated auditory exposure.
Over time, this workflow has become so fragmented and complicated that I sometimes forget my next step altogether, which is why I started wondering whether I could hand this entire process over to an AI agent.
Below is the messy workflow in ordinal numbers
When I study Japanese from my course: I copy instructor-selected example sentences from the textbook. Every sentence should be treated as important learning material.
For each sentence, the system should automatically: generate a detailed explanation suitable for long-term storage generate a clean sentence-only version for listening practice
The two outputs must be different in content: one explanatory (will go into Notion as my knowledge asset) one minimal (will go into Speechify for listening)
I do not want to decide routing or formatting manually. The system should always do both.
Here is what ChatGPT explained (as if I was talking to myself).
At its core, the difference is not about technical difficulty or whether code is involved, but about what problem is being solved.
What I built with n8n is fundamentally an automation system. It takes a specific input, follows a predefined logic, and reliably produces the same outputs every time. n8n already provides the runtime, the orchestration, the integrations, and the error handling. As long as the workflow works, it can run continuously without me needing to think about interfaces, users, or product structure.
Turning that same workflow into a website or an app would not simply be “making it prettier.” It would mean adding an entirely new layer on top of the automation: a user interface for input, a way to manage state and history, possibly user identity and settings, background job handling, and a more explicit data model. None of these are required for the core idea to work, but they become necessary once the goal shifts from “automating a task for myself” to “building a product for interaction.”
In other words, n8n is sufficient when the goal is to ensure a process happens correctly and automatically, while a website or app becomes relevant only when the focus moves toward user experience, control, and scalability. For my current use case where the system always performs the same actions and does not require manual decisions, sticking with n8n is not a limitation, but an appropriate design choice.
This helped me realize that building an AI agent with n8n and building a web or app product are not competing approaches. They operate at different layers. One optimizes for automation and reliability; the other optimizes for interaction and usability.
I also ask Gemini which one is more difficult to master
In 2026, your observation is widely shared among professional developers: vibe coding a website is often more difficult than building an AI agent once you move past the "honeymoon phase" of a simple prototype.
While building an AI agent involves setting up logical workflows (often in visual, structured platforms like n8n), vibe coding a website requires managing a massive, interconnected codebase through natural language, which introduces several unique layers of complexity.
Still so much to learn!