Earning Judgment

There's a quiet argument going around that the most valuable skill left for humans isn't writing code or prompting well, it's judgment. Not the kind you can grade with a rubric, but the kind that shows up when someone looks at a pile of AI-generated output and knows, in their gut, whether it's actually right for the situation.

The logic is straightforward once you sit with it. Agents can now write, test, and ship code at a pace no human matches. But speed at execution just shifts the bottleneck upstream. Someone still has to decide which problem is worth solving in the first place, someone still has to catch the subtle ways a model's confident answer misses the actual goal, and someone still has to carry the work the last mile past where the machine gave up or got lazy.

That's a different skillset than the one most engineering culture has optimized for over the last decade. We've built entire pipelines, from leetcode interviews to unit test coverage, around measuring things that are gradable: did the function pass, did the tests go green, did the PR merge. Taste and judgment resist that kind of scoring. You can't unit-test whether a product decision was wise or whether a summary quietly dropped the one caveat that mattered.

And that's precisely why the argument holds up. Things that are easy to grade are exactly the things agents will eat first, because reinforcement learning loops thrive on clear reward signals. Ambiguous, high-context calls, the kind that require knowing a customer, a codebase's history, or a market's quirks, don't hand themselves over to automation nearly as easily.

So the practical takeaway isn't to panic about agents replacing engineers wholesale. It's to notice which parts of your own work are actually ungradable, and to deliberately get better at them, because those are the parts that don't scale away.