Some of the best lessons I've ever learned came from disagreeing with a colleague. We were both convinced we were right, and we disagreed on principle.
One of those was about a specific database design decision we were making. I was sure the data should be normalized. He was firmly on the side of denormalizing. We ended up somewhere in the middle, balancing performance against keeping duplicated data in sync. There was no right answer, just a tradeoff we had to understand together.
That kind of friction is how engineers build the judgment that makes them senior. AI is very good at removing friction, and a lot of it deserves to go. But a junior engineer who asks AI instead of asking a colleague gets an answer and skips the argument, and the argument is where the judgment comes from.
So the real question for developing senior talent in the age of AI is which friction to keep.
In the House of Software Value Delivery, the foundation is a well-functioning ecosystem, and two of its Attributes are Ease and Learning & Growth. AI is a Tool that creates a lot of Ease. But by optimizing for Ease, we reduce some of the friction that enables Learning & Growth. Balancing the two is a choice individuals and teams have to make on purpose.
Here are three kinds of friction I'd keep for junior engineers.
Friction with the requirements. Give them the questions. They don't need to write every line, but they need to be able to defend their answers to questions like: Who should be authorized to see what data? What load does this need to support? What happens when an integration changes? Then add friction with the design: how does the design meet, or not meet, those answers?
Friction with the work. Some work gets done without AI, the same way you learn arithmetic before anyone hands you a calculator. Universities seem to be taking this approach, and it has a place on teams too.
Friction with a senior engineer. Make them the first reviewer. Junior engineers review agent-written code first, then walk a senior engineer through why it works and what would break it. It's a strong learning tool with a big tradeoff in efficiency: review is already the bottleneck for AI-generated code, and this slows it down further. That slowdown is the investment.
I've used versions of the first two. The third is the one I'd try next, knowing what it costs.
The full framework is on my website; link in the comments.
Which friction would you keep on your team? And what disagreement early in your career taught you the most?

