How I Built Agent (1): The Most Expensive Agent Nobody Has Used Yet

Nobody has used my Agent yet. With a little self-deprecating humour, I sometimes think of it as the most expensive agent nobody needed. But its original purpose was never to win users. I wanted to discover how generative AI works and what I could actually build with it.

I could have just subscribed

I could have paid for an existing service. It would probably have worked sooner, cost less and spared me a great deal of trouble. Unfortunately, it would also have spared me the most interesting part: the trouble itself, and everything I learned from it.

At first, ChatGPT was my guide

I started by talking to ChatGPT. I asked what to do, why things worked the way they did and how the pieces fitted together. I uploaded files to the server myself, while the chat explained the steps and helped me make sense of technical decisions. It was a rather unusual classroom: instead of studying every technology before starting, I learned while trying to build something real.

Perhaps the textbooks should come afterwards

That experience made me wonder whether conversational AI offers a different way to learn. You begin with a problem, ask questions, challenge the answers and test what works. Of course, an AI can deliver outdated or incorrect advice with remarkable confidence. Verification and independent judgement are essential.

Perhaps now is the time for me to read the textbooks I skipped at the beginning. With practical experience behind me, many concepts might finally fall into place. I might also discover that some things could have been done far more simply. But had I studied everything first, I would have missed all those detours, dead ends and unexpected paths. Some were unnecessarily long; all were part of the experience.

Books about rapidly changing AI tools can become outdated quickly. That does not make the fundamentals of software architecture, databases, testing or security obsolete. The useful combination may be to experiment with current tools and then revisit enduring principles.

Then came Codex

Later I moved from chat-guided manual file uploads to working with Codex. The focus shifted towards repositories, describing changes, reviewing results and testing. Greater automation did not remove the need to understand the system. If anything, it made architectural decisions and verification more important.

The experiment grew into a system

Over time, Agent gained public and private components, a knowledge base, interfaces for other applications, analytics and data synchronisation. The architecture brought together Cloudflare Workers, D1, vector search, WordPress, PHP and MariaDB on WEDOS hosting. Integrating them introduced challenges of its own, including access control, resource consumption, backups and recovery.

At one point the database grew towards 1.7 GB during synchronisation work. That was a practical reminder that making a process work is not enough: it also needs limits, cleanup, monitoring and a recovery plan. Some parts of Agent still need further implementation or independent verification.

What did I actually gain?

If success were measured only in users, there would be little to celebrate so far. But that was never the point of this experiment. I am glad I started it. I have gained experience I would not have acquired simply by reading about generative AI.

I also do not see Agent as a finished experiment. I expect it to become part of my wider application ecosystem, helping me manage and coordinate the services I build. That is an ambition, not a claim that everything is already operational.

Even if its future turns out differently, I do not regret the journey. So far, the most valuable result is not the software itself, but what I learned by building it.

Next: What development looked like when ChatGPT guided me and I still uploaded the files to the server by hand.