Judgement is the new talent

We put a team of young developers in charge of code that was older than most of them.

Seven applications, thirty years of patches, billions of dollars in operating revenue running through them, and a service provider that was winding down whether we were ready or not. We had bought the legacy software from them, and responsibility for keeping it running was moving to us.

Naturally, we asked for the documentation.

We got a user manual from 1996.

Most of the original developers had retired. The few who remained knew their corners of the system, but after three decades of fixes and additions even they sometimes had to work things out as they went.

So we built a new support team of junior developers and put a seasoned manager in charge of it. There wasn't much new development involved, because the work was mostly reconfiguration and a great deal of reverse-engineering.

They had to open code written by someone else, sometimes decades earlier, that may have been touched by dozens of developers since. They had to work out what it did, why it was written that way, what else depended on it, and whether it would behave the same way in our environment.

At the time we thought we were teaching them how to support an old system.

As AI takes on more of the production of code, the developer role starts to shift. Writing the code becomes less of the differentiator. Understanding what the code is doing matters, but so does understanding what happens around it: what it depends on, what depends on it, what could go wrong, and whether the result is actually ready for production.

That is very close to what this team spent 11 months doing.

They couldn't assume the code was right because someone else had written it, and they couldn't rely on documentation to tell them. They couldn't understand one function in isolation from the rest of the system, and they couldn't stop at "it runs."

Over time, the question became less about whether a change worked and more about what might break downstream if it did.

They had to develop judgment.

We certainly didn't design that transition as an AI-era developer training program, but it may turn out to have been one.

If you're hiring developers this year, how much are you testing their ability to write code, and how much are you testing their ability to judge code they didn't write?