I tried replacing a website admin area with an AI agent

25 October 2026 · Case studies · 3 min read

In early 2026 I tried something new.

I wanted to step away from WordPress and the other PHP admin areas I'd built in the past and build a website that could be run entirely by an AI agent working on my behalf. I had recently read about how AIs were now "agentic" - they could run tools more easily and actually perform actions rather than just answer questions.

I also recently met with a consultant who had explained he thought AI agents would be at the core of all business software in the future. Users would no longer manually enter information on screens and forms - they would instead converse with an AI agent who would input the data into the system for them. Thinking about this myself, I started to see how he could be right for some software, like CRMs and basic data entry. Privacy concerns aside, it could make data management tasks much easier for businesses.

The plan

My plan was to build a hiking adventure website and have an AI in the middle orchestrating everything. I'd talk to the AI in the browser, then it would help me put posts together and write the content with me. I could have it add the post to the database and deal with images and metadata.

It started so well

I used Claude to help build this. We started by designing the front-end, which is shockingly easy on Sonnet or Opus if you know what you want. Claude built the bare-bones in less than 30 minutes.

We then started to work on the AI agent. I gave Claude the instructions of what I wanted. The clear parts around being able to create, read, update and delete posts, images, areas and points of interest. I also wanted a way of uploading GPX files and displaying them on the post pages.

Why didn't I like it?

I started using the agent and the first thing I noticed was the speed - it was very slow - much slower than I was expecting. The model was set to OpenAI's GPT-5-Pro. It was smart but it took over a minute to think then run the tools, sometimes it even hit the PHP timeout. I needed something a little faster. I read up on how others had been doing this and they recommended using a nano model instead for the orchestrator then give it a writing tool using GPT-5-Pro for writing the post content.

Switching to the GPT-5-Nano model fixed the speed problem but introduced other issues. The main one was hallucinations and mistakes. I would ask the agent to create a new hike post and add some images in, then if I tried asking it to edit the post it would seem to start fabricating content and bringing in facts from other posts. I managed to fix some of this by reducing the amount of memory the chat had. It seemed to be remembering things in the chat and mixing them all up.

Another small pet peeve I noticed was the tools not being quite perfect for what I wanted the AI orchestrator to do. I managed to get the AI to tell me what was wrong then I'd paste the message to Claude and get the issue fixed. This didn't line up with my vision of the AI running the whole website if it needed outside help from another model.

What I'm going to do instead

When the faster AI models get smarter I'll be having another try at this. At the moment I'm going to use Claude with both an MCP server connection and an SSH connection to the website files. This will give Claude access to run the tools we've already built, like being able to manage posts and upload images, plus with the SSH connection it will be able to make any changes to the HTML files and even make changes to its own tools if needed.

I find Claude's thinking and agentic loop is much better than anything I could build with an OpenAI model at this time. I do expect things to change in the future, and I'll be revisiting this as models evolve.