They are hiring builders. The work is supervision.

The job posting and the job are different things.

At the end of a long day, in a room of about 80 leaders I taught this month, people filled out a one-page plan for the first thing they wanted a machine to do for them. One of the fields asked what "working" would look like after 90 days.

Most of the answers were modest. A report that runs on its own. A dashboard that stays current. A file that stops needing a person.

One answer was more ambitious than the rest, and it is the one I keep thinking about:

Read that again. It asks for something specific:

That is not a machine that replaces the person. That is a machine that earns the person's inattention.

I would call it the most sophisticated request in the room, and it was written by someone who was not trying to be sophisticated. They were trying to describe their Tuesday.

Now consider what the market is doing.

Before the session, I sent the same room a pre-work survey with a simpler question: have you ever set up anything that runs on its own? Anything. A rule, a scheduled report, an automation, a filter that does its job while you sleep.

47% said yes.

And here is what those same people are walking back into.

Start with the person applying for the job at the bottom of your org chart. Indeed's Hiring Lab published two reports on July 23. The first found that in May, 30% of applications to entry-level roles came from people with 10 or more years of experience, more than any other group. Nearly half of all applications from that group went to entry-level jobs.

The bottom rung is crowded with people who climbed off it years ago.

The second found senior-level postings up 14.7% year over year and entry-level postings down 7.5%, both as of May. In software development, senior roles were 69.3% of postings in the first quarter of this year, against 4.5% entry-level.

And the bar for getting onto the bottom rung is rising. In NACE's Job Outlook 2026 Spring Update, fielded in February and March, employers said more than a third of entry-level jobs now require AI skills, nearly triple what they said in the fall. That is 185 employers answering a survey, so read it as a direction rather than a census. Handshake's Class of 2026 report, from 1,248 students across nearly 500 institutions, found that 28% of seniors say AI has been meaningfully integrated into their academic program.

I am not going to tell you AI did this. The New York Fed's research director published a note on August 5 citing very few AI-driven layoffs, and concluding that the effect so far runs through changing skill requirements rather than eliminated jobs. That is the claim that matters: the entry point moved.

So the market is compressing the bottom rung, raising the bar for what a new person is expected to arrive already knowing, and describing that bar in the language of autonomy.

Build the agent. Run the stack. Ship the thing that ships itself.

That is not a caricature. Two postings live right now, neither of them an engineering role. GitLab wants a Revenue Technology Analyst who works “under the direction of senior team members,” and asks for experience “leveraging AI in your day-to-day workflows, ideally building agents and artifacts that have proven to accelerate productivity.” Squarespace wants a Media Strategist who uses AI “not just chat prompts, but building workflows, automations, and agents that make the whole team faster,” and then adds: “you can show us what you’ve built.”

Meanwhile the people who will be on the hook when it breaks are asking for something quieter and harder: a thing they can stop watching, with a promise about when to start watching again.

Those are different jobs. The market is hiring for the one that builds. The work is the one that watches.

I see this pattern in the data I have: the skill that gets screened for is building; the skill the work actually asks for is supervision.

No one writes a job posting for that. Everyone needs it.

──────────────────────────────

→ Half-Baked.
I am not certain the gap I just described is a preference at all. The generous read is that buyers want less than vendors sell, and vendors should meet them there. The less comfortable read is that nobody has built the missing layer yet, so there is no autonomy on offer to decline in the first place. What would make real autonomy survivable is failure detection you can trust and a clear answer to who holds the credentials, and I do not think either is solved.

Two of the most aggressive operators I have watched publicly, people running large volumes of work through agents every day, both built their practice around a human checkpoint. One of them keeps production credentials out of his agents' hands entirely and checks his highest-stakes threads every 25 minutes. If the people pushing hardest are still supervising that closely, then "buyers want less" may be the wrong frame entirely, and the right one is "less is what exists today."

I genuinely don’t know which it is. The distinction matters, because one version says the category is oversold and the other says it is early. I am leaving it open.

→ Hot Take.
"Agentic" is doing the same work "AI-powered" did in 2023. It describes the seller's ambition for the product, not a measured property of the thing that ships. And the cost of that gap does not land on the seller. It lands on whoever bought it expecting the label to be literal.

The version of this that is coming, and I would bet on this: "manages AI agents" is going to become the new "proficient in Microsoft Office" of 2028. A line on a résumé that everyone writes…and no one verifies. And it stops carrying information the moment it becomes universal. The people who can actually prove it will be the ones who can point at something running right now that they no longer watch.

→ Confession.
I built a segment of that teaching day to make an argument about where agents are going: here is what changes when work runs without you standing over it. I was proud of it. It was the piece I most wanted to give them.

Then the pre-work numbers came back, and fewer than half the people who answered had ever set up anything that runs on its own.

I had aimed a segment about supervision at people who had never had anything to supervise.

The segment was good, and the room took something from it. But it was pitched above where they stood, and I did not know that until after I had built it. I designed for the frontier and taught to the middle, and the correction came from a form I had written myself.

──────────────────────────────

Reply with yours. I read every response, and the interesting ones become future issues.

──────────────────────────────

Take the automation you rely on most right now. It does not have to be AI. A rule, a report, anything recurring.

Finish this sentence the way our example did:

"Working would be me not having to ______ unless ______.”

Whatever fills the second blank is your actual specification. It is almost certainly smaller and more specific than anything you have been shown in a demo. Write it down. That sentence is worth more in a vendor conversation than any feature list you could bring.

List the tasks a new person traditionally learns on: the ones you handed someone in their first 90 days because doing them badly was cheap, and doing them at all taught them how the place works. Call it an entry-rung audit.

Then mark which of those a machine now does.

You are not looking for a headcount answer. The rung was never valuable for the tasks. It was valuable because it was the cheapest place in the building to be wrong, and being wrong cheaply is how judgment gets built.

So the real question is not where the next person learns those tasks. It is where being wrong is still cheap. Supervising a machine is one of the few honest answers: catching a bad output before it ships is the lowest price you will ever pay for the same lesson.

The rung did not disappear. It moved.

Assume the management layer is the growth market. Not management of people, which we have had for a century and are reasonably good at. Management of systems that produce work while nobody is watching. We have real practice at this for infrastructure, servers, plants and networks, and almost none of it for knowledge work. The discipline exists. It has not crossed into the office.

Every generation of that work has needed a class of person who can tell the difference between quiet because it is working and quiet because it broke. Quiet is not the same as working. We have not built that class yet, and the people who step into these roles will not be the ones who built the most. They will be the ones who can say precisely what they stopped watching, and why they were right.

──────────────────────────────

The Supervision Spec. Three lines. Fill it in before your next automation conversation, internal or external.

> 1. The thing I would stop checking: ______________________

> 2. The condition that makes me start again: ______________________

> 3. Who finds out when that condition is met, and how: ______________________

That is the whole instrument. Line one is what you are actually buying. Line two is the specification nobody writes down and everybody needs. Line three is the one that gets skipped, and it is the one that decides whether any of this survives a vacation.

If you cannot fill in line two, you are not ready to stop checking. That is not a failure.

That is the finding, and it is a cheaper thing to learn on a piece of paper than in production.

──────────────────────────────

I write this newsletter on a simple bet: that most of what matters about this technology is happening on a curve, not at a cliff, and that the people who read the curve early get to make different decisions than the people who wait for a drop.

In a recent issue I argued that the evidence about whether AI works is almost entirely secondhand, and that your own record is scarce and valuable because of it. This issue is what that looks like at the level of a single job. The market is describing a capability in public. The people doing the work are describing a different one, quietly, on forms no one reads. I have read some of those forms. They are more interesting than the job postings.

If something here changed how you are thinking, hit reply. I will respond.

Thank you for reading. If you know someone who will benefit from this thinking, please share.

— Chris

Columbia Faculty · Chief of AI & Strategy at Essential Innovations · Founder, Attainable AI

chrishuberreitz.com · LinkedIn

──────────────────────────────

Sources

Indeed Hiring Lab, Entry-Level Jobs Aren't Just for Inexperienced Workers, July 23 2026

Indeed Hiring Lab, The Labor Market Is Tilting Toward Seniority, July 23 2026

NACE, Job Outlook 2026 Spring Update, April 20 2026 (185 employers)

Handshake, Class of 2026, April 2026 (1,248 students, ~500 institutions)

Federal Reserve Bank of New York, AI's Impact on Labor and Hiring, Liberty Street Economics, August 5 2026

GitLab, Revenue Technology Analyst, accessed August 19 2026

Squarespace, Media Strategist, accessed August 19 2026