AI Agent - Treat them like an intern

AI Agent - Treat them like an intern.

Like many experienced programmers, I am horrified fairly regularly by the vibe coded slop being pumped out at unprecedented volumes.

With that said I realize the profession of “Programmer” will likely go the way of the elevator operator some day.

It’s “evolve or die”… That is a bit dramatic. Perhaps “evolve or be unemployed”.

That is why I am not completely throwing AI out the window, but I am also NOT delegating my thinking to the LLMs either.

LLMs are great parrots and function fairly well given a linear problem but systems architecture is not a linear problem. Yes, there are many linear processes that flow through a system but the system as a whole is a 4D system that functions in real time.

Changing part of it has rippling effects on the rest of a system, and even the latest models seem to have trouble keeping even a small fraction of the entire system in their context window.

So here where I am at as of Q3 2026 (This might change someday later):

Treat it as an intern that can Google remarkably fast.

I will concede this: Claude, Gemini, etc. can all search Stack Overflow 1000x faster then I can.

So I often use it to search just like I used to Google error messages 10 years ago but I always ask it to give me the source links as it is prone to jumping to conclusions that make no sense.

The other way I have had success is by completely hand-coding a framework, then wiring in the framework to about 4 or 5 places that the AI can use as an example.

Once I get that far I task the agent with using those examples to start wiring in my framework across the system.

It’s not allowed to edit the core code I wrote. If it runs into a limitation or a feature that is required it needs to come back to me and we solve it together.

Many times over the years, this is how I worked with teams. I would design a system, code up the core framework it is all built off of, then hand it to a team of humans to copy my base implementation and wire it in.

The only difference now is that I can do it much faster with agents running 24 hours a day.

I still review every line of code that hits production. Now with even more attention to detail than in the past, I knew and trusted my team. Now it's dependent on a probabilistic mode that "probably" will write code that works…

In the end I end up taking a play from the Dan Martel playbook where I do the first 10%, the framework, the agents end up doing about 80% of the work and then I do the last 10% final code review and touch up.