People in the Agentic Loop
For about a few thousend years, getting good at a job worked like a staircase. You started at the bottom doing the boring, repetitive, slightly humiliating work. You watched people above you. You absorbed a thousand tiny judgments nobody ever wrote down. Eventually you became the person other people watched.
Then we handed the boring, repetitive, slightly humiliating work to a machine that never sleeps, never complains, and is wrong with total confidence roughly fifteen percent of the time.
The workforce is changing, and not gently
A few things broke at the same time, which is why everything feels chaotic rather than merely difficult.
The pyramid is missing its base and no longer needs much of its middle. The junior work that used to justify hiring juniors is now the work the machine does best. So the easy hire to skip is the entry-level one. Which is a bit like saving money by not planting trees, and then being shocked in ten years that there is no shade.
The value of knowledge and the half-life of skills both collapsed. Knowing the answer or even more so - the right questions to ask used to be a career. Now the answer or questions are almost free, instant, and occasionally hallucinated. What you knew last year is depreciating faster than a new car driven off the lot.
The bottleneck moved from execution to judgment. We spent decades optimizing how fast people could do things. We are now drowning in things done fast, and starving for people who can tell which of those things were worth doing.
The curriculum was designed before the technology existed. We are still training people for a race that changed tracks halfway through, and grading them on how well they run on the old one.
These skills cannot be assessed in a sixty-minute test. You cannot multiple-choice your way to good judgment in ambiguity. We have built an entire credentialing industry around measuring the one set of things that no longer needs measuring.
We confused training with credentialing for thirty years. A certificate proves you sat in a room. It says almost nothing about whether you would catch the expensive mistake before it shipped.
And the tacit stuff - the unwritten, learned-by-osmosis, "you just know" stuff - transfers through apprenticeship. Then we cut the apprentices. We removed the exact mechanism by which the irreplaceable knowledge gets passed down, and we did it because it looked like a rounding error on a spreadsheet.
So where do humans fit in the AI loop?
Here is the good news, and it is genuinely good. The things that are hardest to automate turn out to be the things that were always quietly the most valuable. We just never paid for them directly, because they came bundled with the people who also did the easy work.
Critical thinking, verification, and calibrated skepticism. The machine produces a confident paragraph. Someone has to know it is subtly, dangerously wrong. Verification is no longer a chore at the end. It is the job.
Taste. The unglamorous, unteachable sense of what is good. Taste does not scale, cannot be prompted, and is the single hardest thing to fake. Which is exactly why it is about to become extremely well paid.
Ownership and judgment under ambiguity. Anyone can research and decide when the answer is obvious. The rare skill is deciding well when it is not, and then putting your name on it.
Curiosity, and learning how to learn - then teaching others. If skills have a short half-life, the only durable skill is the ability to keep acquiring new ones, and to help the people around you do the same.
Comfort with ambiguity. The willingness to act before the picture is complete, because the picture is not going to be complete again soon.
The human network. Trust between actual people does not automate. The relationship, the reputation, the favor remembered three years later - that is infrastructure no model can index.
A scientific approach. Hypothesis, experiment, evidence. In a world of plausible-sounding output, the person who insists on testing the claim is worth their weight in compute.
High-trust human work, from introverts to extroverts. Not just the charismatic room-workers. Also the quiet person everyone trusts to be right. Both are now premium.
Project management for everyone. Schedule, quality, resources, cost, and risk. When execution is cheap and abundant, orchestrating it well becomes the differentiator. Everyone is a little bit of a project manager now, whether they asked to be or not.
Notice the pattern. None of these fit in a sixty-minute test. All of them used to be the "soft" stuff we treated as a bonus. The soft stuff just became the hard currency.
What we can actually do
Diagnosis is comfortable. Everyone enjoys describing the storm from inside. Here is the less comfortable part - the list of things you do on Monday.
Rewrite job descriptions around what the role decides, not what it does. If the description is a list of tasks, it is a description of something a model will be doing shortly. Describe the decisions, the judgment, the ownership. Hire for those.
Protect the apprenticeship pipeline. This is the one that matters most and gets cut first. The easiest saving today is the most expensive mistake in 2031. You cannot grow senior people without growing junior people, and you cannot microwave experience.
Make AI literacy table-stakes, not a bonus. Knowing how to work with these tools is no longer a special skill worthy of a workshop and a snack. It is the baseline, like email, like typing. Treat it that way.
Measure verification, not just output. Reward the person who caught the error, not only the person who produced the most. If you only measure how much gets shipped, you will get a great deal of critically wrong work shipped very quickly.
Promote on what people kill, not just what they ship. The most valuable judgment in an age of abundant output is the judgment to say no. The work someone chose not to do, the bad idea they stopped, the feature they killed - that is signal, and most companies are blind to it.
Treat skills as reusable internal automations. A thing learned once should become a reusable internal IP through AI skills, agents, etc., not a hero's private knowledge that walks out the door at five o'clock.
Demand that people learn by using AI - set aside the time, and check the results. Everyone has access to a fairly competent expert in almost anything with inexhaustable patience, provided you've got the tokens. Learning by doing, with guardrails. Protect the hours. Then actually look at what came out.
Burn the annual training cycle. Build continuous infrastructure. A once-a-year course about a field that reinvents itself every quarter is theater. Build learning into the work, not into a calendar invite in March.
The point
We are not removing humans from the loop. We are changing what the loop needs them for. The machine handles the execution. The human handles the part that was always the hardest and the most human - deciding what is worth doing, noticing when something is wrong, owning the result, and carrying the trust.
The companies that win the next decade will not be the ones with the most tools. Tools are about to be free and everywhere. The winners will be the ones with the best judgment about when to trust them, and what to do in the first place.
Which is, when you think about it, a very old answer to a very new question. The robots took the homework. The grown-ups still have to decide what the homework was for.