Your AI Agents Are Probably Just Traditional Software

I know that sounds like I’m trying to be provocative, but the more I build this stuff, the more obvious it becomes.

I’ve spent this weekend working on a small company-monitoring agent. The bit everyone would call “AI” is fairly narrow: I’m using Gemini to help retrieve, interpret and classify business information.

That bit matters. It’s useful.

But it is not the whole product.

Around that, I still had to build:

• a PostgreSQL data layer
• Docker orchestration
• MCP tools using FastMCP
• Python logic
• an HTTP wrapper around Companies House
• the execution logic for when the process should run, what it should check, and what happens next

In other words, most of the work was ordinary software engineering.

This is where I think the current language around AI gets messy.

AI is often talked about as if it sits outside the normal software engineering process. As if once there’s a model involved, we’re suddenly in a different world where architecture, state, data modelling, permissions, retries, monitoring, testing and deployment are secondary details.

They are not.

The model is a component.

A powerful component, yes. A very useful one. But still a component.

A lot of “AI agent” work is really:

Prompt → model call → tool/API call → state update → next step → repeat.

That can be valuable. I’m not dismissing it at all.

But the orchestration is software.

The data layer is software.

The integrations are software.

The reliability work is software.

The scheduling is software.

The permissions and audit trail are software.

Only part of the system is AI.

I think this distinction matters because the way we talk about AI is starting to create a strange split in the engineering world.

“AI people” on one side.

“Software engineers” on the other.

That feels wrong to me.

The best AI products I’ve seen are not built by treating AI as magic, or as something separate from engineering. They’re built by people who understand how to put a model inside a well-designed system.

The hype will settle down eventually. It always does.

And when it does, I think we’ll be left with a much clearer view:

AI is not replacing software engineering.

It is becoming part of it.