One-Third of Entry-Level Jobs Now Want AI Skills. The Edge Is Managing Agents, Not Using Them. | Edition 304

For two years, the entry-level story ran in one direction. AI does the junior work, so companies stop hiring juniors. Then in May, The Wall Street Journal ran the opposite headline: companies saying AI is reviving entry-level jobs, not killing them.

Both versions have evidence behind them, and the reason they can coexist is the genuinely useful part of this story. The entry-level work that vanished and the entry-level work that appeared are not the same work.

The sharpest evidence sits in neither headline. It is in a Harvard Business Review article from February that put a name to a job a lot of people were already half-doing — and in a set of hiring numbers that show precisely where the leverage moved.

Harvard Business Review named the job in February

On February 12, 2026, Harvard Business Review published To Thrive in the AI Era, Companies Need Agent Managers, by Suraj Srinivasan, a professor of business administration at Harvard Business School, and Vivienne Wei of Salesforce.

Their definition is deliberately unglamorous. Agent managers are “leaders responsible for orchestrating how AI agents learn, collaborate, perform, and work safely alongside humans.”

The comparison they draw is to the product manager — a role that did not exist in recognizable form before software ate everything. It was invented because somebody had to stand between what engineering could build and what the business actually needed. The argument is that autonomous agents open the same gap, and it needs the same kind of person standing in it.

Notice what is not in that definition: writing code, training models, or knowing how a transformer works. It is a specification-and-review job. You decide what the agent should do, you judge whether the output is good, and you own the exceptions it cannot handle.

The demand shows up in the survey data

The National Association of Colleges and Employers runs one of the larger recurring surveys of employer hiring plans for new graduates. Its Job Outlook 2026 Spring Update — fielded February 12 to March 17, 2026 — found that more than one-third of entry-level jobs now require AI skills, according to the employers themselves. That is nearly triple the share who said so in fall 2025.

Nearly triple, in roughly six months. Whatever else is true about the graduate job market, “AI skills” moved from a differentiator to a baseline expectation faster than almost any credential in recent memory.

Gartner's projection points the same direction over a longer horizon: by 2029 it expects at least half of knowledge workers to develop new skills to work with, govern and create AI agents on demand for complex tasks. That is a forecast rather than a measurement, and worth weighing accordingly — but it is a forecast about governing agents, not about using chatbots.

The part that should temper the optimism

Here is the number that keeps this edition honest.

Indeed's Hiring Lab published an analysis on July 8, 2026 by economist Guillermo Gallacher. US software development job postings have grown by almost 15% since the launch of Claude Code in late February 2025, while overall job postings fell 7% over the same period. That is a real reversal — the roles most exposed to AI are the ones that rebounded hardest.

But 71% of the increase in software development job postings between May 2025 and May 2026 came from senior roles.

So the demand is real and it is lopsided. Postings came back; they came back mostly above the entry line. Anyone telling you the entry-level squeeze is over is reading half the data. Anyone telling you AI has ended entry-level work is reading the other half.

The defensible reading is narrower and more useful than either: the tasks that used to justify a junior hire are increasingly done by agents, and the thing still worth paying a person for is the judgment layer sitting on top of them.

Why managing agents is a different skill from using AI

That distinction sounds like semantics until you look at what each one actually involves.

Using AI is a single-turn transaction. You have a task, you ask, you get an output, you use it or you do not. The skill ceiling is phrasing the request, and it is low — which is exactly why it commoditized so fast.

Managing agents is supervisory work with the same shape as managing people, minus the feelings. You have to decompose work into pieces small enough to hand off. You have to write a specification precise enough that success is checkable. You have to build a review step, because you are now accountable for output you did not personally produce. And you have to know which work should never leave your desk at all.

Anthropic's Economic Index, which measures how Claude is actually used rather than how people say they use it, splits observed conversations almost evenly: as of its May 2026 period, 51.4% augmentation against 48.6% automation. Roughly half of real usage already looks like delegate-and-check rather than ask-and-answer. That describes observed usage of one assistant rather than the labor market — but it is a measurement, not a survey of intentions.

A July 2026 HBR research piece by Jim Doucette and Vishal Gaur put the employer side of it plainly: generative AI is transforming what employers expect from knowledge workers not by replacing expertise but by raising the bar for it. You cannot review what you do not understand. An agent manager who cannot evaluate the work is just a forwarding address.

What to do this week

The practical version of this is smaller than the trend pieces suggest.

• Pick one recurring task and write a real specification for it. Not a prompt — a spec: what the input is, what a good output looks like, what the failure cases are, and how you will check. If you cannot write the check, you cannot safely delegate the task, and that is worth knowing about your own work.
• Run it three times and review it like a manager, not a user. The question is not “is this good?” It is “would I sign my name to this without reading every line?” Wherever the answer is no, that gap is the job.
• Keep a list of what you would never hand over. Anyone can list what AI does well. The person who can say precisely where it should not be trusted, and why, is describing judgment — and judgment is the part of the work that is not being posted away to senior roles.

If you are early in a career right now, the honest summary is this. The market did not get easier; the Indeed data says the recovery landed above you. But the skill that closes that gap is not “learn AI” in the generic sense that every job posting now lists. It is the narrow, teachable, demonstrable ability to hand work to a system, verify what comes back, and own the result. That is a management skill — and it is the first one in a long time that you can practice before anyone gives you the title.