The winning AI team can't grow its own seniors

Anyone watching entry-level hiring fall gets handed the same reassurance. It’s the market, the story goes, and it’ll correct; upskilling covers the gap until it does. Junior developer hiring is down 35 percent, bootcamp enrollment is down 40 percent, and the comforting version of the story says this is a dip, not a demolition.

Now set that next to a number McKinsey published this year. In its report on the “agentic organization,” the firm found that two to five people can supervise fifty to a hundred AI agents running an entire process end to end, like customer onboarding or a financial close. AI agents here are software that carries out multi-step work on its own, not chatbots that answer one question and stop. McKinsey calls the ratio a credible target for any process with repeatable steps and clear success criteria, and says it tested with early adopters.

Read the two facts together and the reassurance falls apart. The cause is sitting in the org chart everyone’s racing to copy.

The lean AI team that works today deletes the exact seat where its own seniors were made, and it runs at all only because it’s staffed by people the pre-AI world trained.

Look at what the fifty to a hundred agents actually do. McKinsey is specific: they handle the execution. The two to five humans do coordination, exception handling, and judgment escalation. That’s senior work by definition. And the execution the agents absorbed is the exact work that used to be the apprenticeship. Writing the boilerplate. Building the unit tests. Keeping the documentation current. A separate strand of the same research, on how the best teams reorganize, names the identical list: high performers push what it calls glue work (documentation, test generation, code-review support) onto AI so humans can concentrate on judgment [reported]. That glue work was never only grunt work. It was how a junior learned the domain by shipping in it.

So the AI-native team trends toward three to five senior engineers where eight to twelve people used to sit [verified]. Early-career roles contract as AI takes the boilerplate [verified]. And here’s the tell most write-ups skip past: teams that flattened their seniority to save money have generally regretted it inside two release cycles [reported]. They need the seniors. They just removed the place seniors come from.

The usual framing calls this an unsolved tension, as if a piece were missing and a training program could supply it. But nothing is missing. The org form causes the shortage it then suffers from. A company adopts the two-to-five ratio because it’s cheaper: drop the execution seats, keep a few seniors, add agents. That single move removes the only place seniors were ever grown. As more of the field settles on the winning shape, it burns faster through a supply of senior judgment that nothing is producing. Success is the drain.

The honest objection is that seniors don’t need to have done the grunt work to supervise it. Maybe new pathways manufacture the judgment without the apprenticeship: simulations, AI-tutored onboarding, training people straight into oversight. And the labor numbers look roomy. One projection has AI creating a net 78 million roles by 2030 [reported]. Surely the rung reappears somewhere.

Take the pathways first. The judgment those two-to-five supervisors sell is knowing when an agent’s output is wrong and which exception to escalate. That isn’t abstract. It’s pattern recognition built from having handled the normal case thousands of times. You can’t tutor someone into spotting the anomalous transaction if they’ve never processed ten thousand ordinary ones. If agents do all the execution, no human banks the reps the judgment is built from. And the hopeful pathway isn’t just unproven. Only 2.5 percent of AI engineering roles target people with zero to two years of experience [reported], and retraining access isn’t expanding. The door isn’t widening. If anything it’s narrower.

The 78 million is a headcount claim, not a rung claim. New roles are appearing at and above the skill floor, not below it. The market is raising the floor, not building a ramp. A million new senior-shaped jobs rebuild an apprenticeship about as well as adding penthouses rebuilds a demolished staircase.

This also moves the problem to a more useful place. Addy Osmani, a widely read engineering writer, has warned about knowledge debt: juniors who skip the foundations become brittle seniors. He frames it as a property of people. The ratio makes it a property of org design. The winning structure has no foundational layer for anyone to skip, because the foundation got handed to agents that will never graduate. The fix, if there is one, is to keep some uneconomic execution seats on purpose, as a training cost. That’s exactly the seat a cost-driven two-to-five adoption is built to cut.

I want to be careful about the leap. The numbers are solid. The ratio, the junior contraction, the seniority inversion are all measured. The claim that supervision judgment can only be grown from execution reps, and therefore that this form can’t refill its own senior stock, is my read of the mechanism, not a measured result. It deserves a test, and the test is clean.

So here’s the question to carry into any lean AI team you stand up. Is there a seat in it that a person can enter without already being senior, and grow into one of the two-to-five? If the honest answer is that you hire your supervisors already made, from companies that still have the seats you cut, then you don’t have a steady state. You have a company drawing down a shared reservoir that every company copying the same form is drawing down at the same time, and no one is refilling.

That’s not a reason to abandon the ratio. It works, and today it delivers. It’s a reason to price the thing the velocity numbers hide. Go find one org running this form that keeps a real entry seat and can show you someone promoted through it into supervision. If it exists, the drain is solvable and I’m wrong in the best possible way. If it doesn’t yet, the seat is yours to build, and building it now is cheaper than discovering in ten years that nobody can.