Late Septemebr 2024 was my real entry point into AI agents. Discovering Cursor, understanding what agents could actually do — it was equal parts exciting and unsettling.
Two questions sat with me from the start:
- Can I trust the code these tools generate?
- Can I scale this into something real?
And here's what I was not worried about — being replaced. I never bought into that fear. Because the best programmers were never the ones who memorized syntax. They were consultants. People who know which knowledge to pull from, how to combine it, and what to build with it. That judgment doesn't live in a model. AI can accelerate the work — but thinking, deciding, and growing a product? That's still yours.
Change management was never really about the solution. It was always about people — how they adapt, how they evolve.
So when the agent went through my codebase and started changing things, I made every mistake in the book. Wrong prompts. Wrong actions. No testing. Over-relying on the agent. Fixing loops that never ended. Things that didn't work, and when they did work — they weren't beautiful.
Then Claude Code came out with plan mode. I stopped chasing other tools and focused on building the right way.
Here's what I know now:
AI Agents are excellent at moving you from A → B → C … all the way to Z.
Never ask them to go A → Z on their own.
They don't carry the full memory of your project. They'll repeat themselves, contradict themselves, and slowly hollow out the part of the process that you — as a builder — actually love. The thinking. The architecture. The craft.
Use them as a skilled collaborator on each step. Not as the one holding the map.