I finally stopped treating AI agents like fancy chatbots

Tony Iverson

Sep 3 • Builds & Wins 🚀

I think a lot of developers are stuck in this weird middle ground with AI.

We know how to prompt ChatGPT.

We’ve played with Claude Code.

We’ve heard about MCPs.

We’ve probably built a few little agent experiments.

But getting from “this is cool” to “this agent reliably does work in my actual codebase” is a completely different problem.

What helped me was thinking less about individual AI tools and more about the system around them.

For example:

Claude Code → actually works against the repo

MCPs → gives agents access to the tools/context they need

Skills → makes repeatable tasks more structured

Orchestration → gets multiple pieces working together instead of having one agent try to do everything

That last part is where it gets interesting.

Instead of asking an agent to magically build an entire application, you can start giving it smaller responsibilities and connecting those pieces into a workflow.

I've also found that keeping the work itself organized matters more than I expected. Floment has been useful for tracking the tasks and moving parts around an AI-assisted project while I’m actually building.

The goal isn't “use more AI.”

It’s getting to the point where you can say:

“Here’s the problem. Here’s my repo. Go do the work.”

…and actually trust the result enough to review, test, and ship it.

How far along are you with agents right now, still experimenting, or do you already have one doing real work in your projects?

0 comments

I finally stopped treating AI agents like fancy chatbots