My coding agents remind me of the young developers I used to manage.
You explain a new project and before the conversation is even finished, they are almost running to the keyboard.
Backend, frontend, database, APIs. Everything starts moving at once.
They want to code. To build. To see things working.
I saw this many times when I was leading development teams, and I still see it sometimes with my students today.
The architecture discussion feels boring. Asking more questions about what the stakeholder actually needs feels like slowing things down. Making a plan can feel like wasted time when you already know how to start building.
Until weeks later you discover that you built a very impressive sports car, while the customer actually needed a tractor.
It takes time to teach young developers to resist that urge to execute immediately.
First understand the problem. Ask questions. Challenge assumptions. Think about architecture, security, dependencies and what could go wrong. Then start building.
What I find fascinating is that I am now doing something very similar with coding agents.
Give an agent a task and it is perfectly happy to start executing. Fast.
And because that execution is becoming so powerful, a wrong assumption can turn into a lot of perfectly working code very quickly.
So I find myself teaching my agents many of the same disciplines I used to teach young engineers.
Don’t start yet. Understand the context first. Follow the architecture. Respect the boundaries. Test your assumptions. Stop and ask when something is unclear.
Of course, the way you teach an AI agent is completely different.
But the engineering lesson behind it feels remarkably familiar.
After almost 30 years in software, I am again spending a lot of my time teaching very capable builders when not to build yet.