Andrew doesn’t exist. But his story has happened millions of times.

On his first day at work, he asked me seventeen questions.

Where could he find the data? Who should receive the document? Was he allowed to call the client directly? Did he need my approval before making a decision?

He wasn’t insecure. He was young.

For the first few months, we worked side by side. I told him what to do next, and he did it. Whenever he ran into a problem, he stopped and came back to me.

Then one day, just before boarding a flight, I gave him an important assignment.

“By tomorrow morning, the client needs everything required to make a decision.”

The Wi-Fi didn’t work during the flight. For nine hours, I couldn’t answer Andrew’s questions, review his choices, or give him permission to proceed.

When I landed, I opened my phone expecting a long list of messages.

There was only one.

Andrew had discovered that some of the data was wrong. He had reconstructed the numbers, developed several alternatives, compared them, and selected the one that best protected the client’s most important goal. Then he asked a colleague to double-check the calculations and sent the document.

He wasn’t asking me what to do.

He was telling me what he had done.

That day, Andrew changed jobs. He was no longer someone who followed my instructions. He had become responsible for an outcome.

But I had changed jobs too.

My role was no longer to give him turn-by-turn directions. It was to define the destination, give him the tools he needed, and establish the boundaries within which he could make decisions.

That day, Andrew and I both changed jobs without realizing it.

It happened again

A few weeks ago, the same thing happened with Loop in Codex and Goal in Claude Code.

Once again, I didn’t notice it right away.

Until then, I had used AI as a copilot. I made a request, the AI responded, I reviewed the answer, and I decided what should happen next.

If I stopped, it stopped.

Loop and Goal change that relationship. Instead of giving the machine a single task, we give it an objective.

The agent can make a plan, take action, examine the result, realize that something went wrong, adjust its strategy, and try again. It can continue working for hours while we are doing something else—or while we are asleep.

OpenAI has described Codex runs that work on a single task for more than six hours without a person constantly supervising them. OpenAI

The old relationship looked like this:

request → response → human review

The new relationship looks like this:

objective → action → evaluation → correction → result

AI hasn’t merely learned to do more things.

It has learned to keep going without us.

From prompting to managing

For years, we studied prompting. We learned to write precise instructions, provide examples, and specify exactly what the answer should look like.

That was the language I used with Andrew on his first day:

Do this. Then do that. Stop when you’re finished and show me your work.

An autonomous agent requires a different language:

This is the objective. This is the context. These are the boundaries. This is how success will be measured. If one approach fails, try another. Call me only when you encounter a decision you cannot safely make on your own.

This is no longer just prompting.

It is management.

The best manager of an AI agent is not necessarily the person who can write the cleverest prompt. It is the person who can choose the right objective, transfer the right context, grant the right amount of autonomy, and recognize a wrong answer even when it sounds convincing.

And that creates a paradox.

A technology for old hands

Throughout the digital age, we have assumed that every new technology would favor the young.

Children learned to use computers before their parents. Teenagers explored the internet while adults struggled to understand it. Then smartphones and social media arrived, and once again the young appeared to belong to the future while everyone else tried to catch up.

AI agents may reverse that pattern.

Perhaps for the first time since the beginning of the digital revolution, a new technology may give older workers a greater advantage than younger ones.

Not because older people are better at using the interface. They probably aren’t.

But they already possess what is needed to manage an agent: experience, context, and judgment.

A veteran professional knows which goals are worth pursuing. She has seen enough failures to know what can go wrong. She can distinguish an answer that merely sounds plausible from one that is actually correct. She can delegate the execution to several agents and devote her own time to the decisions that matter.

A young person may learn to operate an AI system in an afternoon.

But you cannot download thirty years of experience.

The problem is that those thirty years used to begin with Andrew’s job.

Today’s senior professionals learned by producing flawed analyses, revising bad drafts, writing simple code, watching experienced colleagues, and making low-stakes mistakes.

Those are precisely the tasks we are handing to AI first.

AI may not merely replace Andrew.

It may replace the process through which Andrew would have become an expert.

Are we removing the bottom rungs?

A career has traditionally worked like a ladder.

People started on the bottom rungs, doing simple work with limited responsibility. Over time, they acquired knowledge and judgment. Eventually, they climbed high enough to see the larger picture and direct the work of others.

AI may not destroy the ladder.

But it could remove the bottom rungs.

Why hire an inexperienced analyst to prepare a report if a senior manager can delegate it to an agent? Why assign routine work to a junior developer if one experienced engineer can manage five AI agents at the same time? Why spend months teaching Andrew to perform a task the machine can already complete?

For each individual company, the decision may make perfect sense.

But if every company makes the same decision, where will tomorrow’s experts come from?

We could create an economy in which people who already have experience become dramatically more productive, while people who still need to acquire it cannot find a place to begin.

A country for old men.

And perhaps no country for young workers.

Of course, the opposite may happen. AI may allow young people to tackle problems that once required decades of preparation. It may not eliminate apprenticeship but compress it. A twenty-year-old with a team of agents may be able to build things that once required an entire company.

We don’t know yet.

But productivity and experience are not the same thing.

Producing an analysis does not mean learning how to analyze. Getting the right answer does not mean learning how to recognize the wrong ones. And avoiding mistakes is not always the same as acquiring judgment.

AI has stopped being our copilot.

We have become its managers.

But if we stop hiring the Andrews of the world, we are left with a question that no Loop and no Goal can answer for us:

Who will become tomorrow’s managers?