Nobody asked you to know that
The part of your job that has a specification is the part that's already gone. What's left runs on things nobody asked you to learn.
Someone put this to me recently and I haven't been able to drop it.
How do you justify, today, with what already exists and not in some projected future, that your job can't be done by an agent?
It took me longer than I'd like to admit to land on an answer I actually believed, and it starts with a concession. A specific one, because a vague concession buys you nothing.
Most of what filled my week five or ten years ago is gone. Working out whether a correction applied to a given system. Finding where a piece of logic was buried inside twenty years of custom code. Reading a dump and knowing roughly where to look. Estimating effort by comparing against a hundred similar cases. All of it done faster and more consistently than I did it, and all of it was on my CV.
What's left has one thing in common.
Everything that got automated was the part that had a specification. What didn't get automated is the part where the specification is wrong.
Why the specification is wrong
Someone asks for a report. The request is clear, well written, and an agent builds it in an afternoon. It builds a good one.
The work was never building it. The work is noticing that the report is being requested because a master data process broke three steps upstream, that building it will quietly make the breakage permanent, and that the right move is to not build it and go have an uncomfortable conversation with the person who asked.
So why does one person in that meeting see it and four don't?
Not intelligence. The four may well be better at their stated jobs. They missed it because the signal wasn't in their lane. To notice that a request is a symptom you need to know something about how that data gets created, who touches it, what the warehouse actually does at six in the morning, why finance needs it that way during close. None of that is in anyone's job description.
The specification is wrong because the people executing it only know the thing they were asked to know.
Which is why the certifications don't save you
A certification is vertical. It makes you more precisely the thing you already are, better at executing the specification you were handed. Valuable, right up to the moment the thing you already are turns out to have a specification. Then it's the worst possible place to have spent three years.
I'm not telling anyone to stop certifying. I'm asking what the learning is for. Almost everything sanctioned, scheduled and paid for by an employer deepens the silo. The kind that widens the field of view is the kind you do on your own, for no immediate reason, and it's the only kind that produces what I described above.
This is also the part of the Da Vinci cliché that's actually worth anything. Not that he had many interests, plenty of people have many interests and produce nothing. That the anatomy made the paintings better. The knowledge crossed over.
Which brings me to augmenting yourself with AI, a phrase everyone uses and nobody defines. Here's my version, and it's narrower than the slogan. If you use these tools to do your existing job faster, you are getting faster at the part that has a specification. That is sprinting toward the cliff. Augmenting means going where you previously had no business being: the finance side of a process you only ever touched technically, the commercial model of the product your system supports, enough of a domain you were never hired for to ask one question nobody else in the room can ask.
That used to take a mentor and two years. It now takes a weekend of genuinely wanting to know. So the constraint stopped being access.
The constraint is appetite, which is much harder to fake than a certificate.
And it explains the person who says this is how it's done here. That sentence isn't said by fools. It's usually said by whoever was right about it for fifteen years and got rewarded every time. They optimised correctly for the market they were in. That's the problem: closed, perfectly executed mastery of a defined scope is now a commodity with a shrinking shelf life, because defined scope is exactly what a model eats. They don't disappear for being stubborn. They disappear for having done everything they were told to do (and nothing more than that).
The move everyone reaches for
The standard answer to all of this is to go into management. Get out of the execution layer before it gets thin.
It isn't wrong. It's that most people reach for it expecting the wrong half.
A great deal of what sits under the word manager is specification. Consolidating status. Reporting upward. Tracking who is on what. Turning six updates into one paragraph somebody skims. That half is being absorbed right now, and the people predicting managers disappear are correct about exactly that half, the same way they were correct about mine.
Which is not new. Years ago the prediction was that Basis would disappear, and developers were the untouchable ones. The forecast has moved onto the developers, and now it comes from inside the vendor: Christian Klein told the Australian Financial Review in June there's a chance nobody is developing software inside SAP in three to four years.
That's the half that travelled. In the same interview he said headcount doesn't necessarily fall, the shape changes, and what the company needs is product managers who can vibe code and who can actually understand businesses. Not someone who codes, not someone who knows the business. Someone whose knowledge crosses over.
The vendor's own forecast, delivered as a warning, describes the person this whole piece is about.
What doesn't get absorbed in that job is harder to put on a CV. Deciding what the team will not do and holding the line when someone two levels up pushes. Knowing which of your people to put on which problem, which requires knowing them and not their skill matrix. Absorbing ambiguity from above instead of passing it down intact, which is most of the job and almost none of the job description.
And the thing usually called empathy, which deserves a precise definition because the greeting card version is useless. Operationally it's accurate modelling of other people's constraints. Knowing a proposal dies framed as a cost and survives framed as a risk. Knowing who needs the decision in writing two days early or nothing moves.
That isn't warmth. It's the same peripheral vision pointed at people instead of systems, and it accumulates the same way.
I don't know how long any of this holds. I've been wrong once already: five years ago I'd have sworn the applicability analysis was pure judgement, and it turned out to be a report. Some of what I'm calling peripheral vision today is probably interpolation over a large sample, and I'll find out which parts the same way I found out last time.
So the answer isn't that an agent can't do my job. It's that the part which isn't specifiable runs on things I was never asked to learn, and the only way to keep it alive is to keep learning things nobody is asking me to learn.
Two questions, and I want the real answer rather than the interview answer.
The last thing you learned properly: was it something your employer asked you to learn, or something you went after because you wanted to understand how it worked?
And if it's been a while since the second kind, what exactly are you betting on?
Pablo
P.S. One thing out of this: pick something adjacent to your work you've been vaguely curious about and never had a reason to study. Not a certification. Something with no credential at the end. Give it a weekend.